Reading their modified license terms, it cracks me up, because they've basically remade the MIT to be the MIT + the one clause that the BSD used to have, which didn't care about MAU or revenue, if you used it in a product, they asked you to 'advertise' them basically. Honestly, its a reasonable request.
Shaming others when all AI is trained off scraped content and code huh? Many of those sources either breaking ToS or being illegal, such as Anna’s Archive. Bold move. And Chinese models in particular have been accused of distilling off American models.
To be clear, the “advertising” clause just requires you to disclose that you use the thing somewhere in the product, such as credits in an “About” section.
I all it advertising clause, because I remember still in the 2000s seeing an Apple ad which at the end of it showed "Unix" or something like that on it, and I remembered that was one of the BSD license requirements, or maybe Apple just did it also just to proudly boast using Unix.
Personally, when I use open code or routers, I feel that beyond a certain level, the models don't make a huge difference to me. Except for expensive and mediocre models like Gemini. In that sense, Chinese models are pretty good. I usually write code in function or method units and then design and assemble them together.
GPT series models are more thorough and better, but I'm not sure if the difference is enormous. It seems to depend on the workflow, but in my opinion, if you are thorough enough, I wonder if there really is a big difference
I've kind of given up on the routers for "free" inference, as you would expect, they tend to give you sub-par thinking because they are obviously trying to conserve as much inference as possible.
I've had some success turning my macbook M1 pro into a heating pad with Qwen 3.6 35B A3B MTP. Trying to use Gemini models "locally" resulted in a similar "short shrift" of effort resulting in mistakes and lots of turns. The reports of Fable being relentlessly "proactive" shows you can go the other direction as well, if you have strong enough branding and effective invoicing.
> I've kind of given up on the routers for "free" inference, as you would expect, they tend to give you sub-par thinking because they are obviously trying to conserve as much inference as possible.
Tangent: did the MTP help you at all? I’ve tested that model back to back on my M1 Max MBP and the MTP version was actually marginally worse. I wonder if I didn’t use the right settings, although I tried several based on the obvious sources.
The difference in outcome isn't that big but yes, you need to be more rigorous. For instance I've found that the Kimi K2.5 and K2.6 models will comment out failing tests rather than fix a problem they just caused (mistaking them for "pre-existing failures"), so you need to specifically make commented-out tests break the build. I've not personally had that problem with any of the Anthropic or OpenAI models.
I wonder why it's the natural tendency of models to BS or do stuff like this when they don't have the correct answer - it's clear that they can program refusal into them, but for some reason, refusal has to be injected after the fact, and models can't really arrive at the conclusion that they can't answer properly.
I really hope we stop using the term "Chinese models". It has this air of Negative connotation. It's the equivalent of calling cars Japanese, which people used to do but now is almost entirely meaningless. You just call them Toyota, Honda, Lexus etc.
I don't think "Chinese" is pejorative in this context any more than "American" is. They are one of the two ecosystems. What's wrong with saying "Japanese cars" today?
For me, it has a positive connotation! In my experience, Chinese Model means cheaper, but still quite effective model you can use for millions of tokens without burning your entire wallet in seconds. That's why I get more excited over a Chinese model release over American models.
I don't know, I tried using one of the Chinese models and it was VERY quick to scan my entire home dir, so maybe your threat surface is a little different than mine
I just had Kimi K2.7-code rebase my Fil-C OpenSSL patch from 3.3.1 to 3.5.7 with quite bare bones instructions and it seems to have worked.
177KB patch, so it's not a small change. The patch did not apply cleanly initially; the agent had to do nontrivial work.
I just showed it the patch against 3.3.1, what command to use to build, and the path to 3.5.7 along with a link to the documentation of the change (https://fil-c.org/constant_time_crypto).
Note, I use my own coding agent (T800, which isn't public, and was previously well tested and tuned for K2.5).
I think this cost me between $5 and $10 in API usage.
I would really love to know if anyone has any experience with something like opencode + Kimi K2.6/2.7 now compared to Claude Code. What is better, what is worse, what is the cost comparison. I am currently paying $100 for the 5x Max plan, but Fable is running through the usage limits quite drastically and I cannot really say it's night and day compared to Opus. Also, I use this mostly for my side projects, so the $100 bill is quite noticeable. I definitely don't want to pay more.
I do have this experience. I've used Claude Code (with Opus mostly), and then switched to opencode (mostly with Kimi 2.6) for my personal projects; it's based on a couple months of use.
Claude Code is better. But Opencode + kimi 2.6 is workable, which is big. For bare code writing, if you know what exactly you want, most popular models are fine (deepseek, kimi, etc), it feels more or less the same as anthropic models.
At the same time, Opus seems to understand my intent way better than e.g. deepseek. I need to be much more precise with my prompts when using deepseek - it often goes in a wrong direction if I'm lazy. This results in a workflow which feels quite a lot different from Claude Code.
Kimi is in between - for me it brings back "lazy prompting" workflow, and I can trust its plans more than deepseek. It enables a workflow similar to Claude Code, it's workable, but it is a bit worse everywhere. Smaller context, a bit more errors, decisions are a bit worse, recommendations are a bit worse, debugging capabilities are a bit worse, etc.
On the usage side, $100 Claude plan is a great value actually. On paper, per-token kimi is way cheaper, but Claude subscriptions are heavily subsidized - you get much more tokens than $100 can buy you. So, in the end, opencode + kimi vs claude code could be of a similar cost, for similar usage patterns. Deepseek can be cheaper, and it has insanely cheap cached tokens, but experience may vary - depending on your habits, you may need to adjust how you work, coming from claude code.
I'd say for side projects something like $10 Opencode Go plan + $10 of extra DeepSeek v4 credits (e.g. on OpenRouter) can be very workable.
This is generally been my experience as well, but i think the main reason for claude code being better at understanding intent is their massive system prompt.
>At the same time, Opus seems to understand my intent way better than e.g. deepseek. I need to be much more precise with my prompts when using deepseek - it often goes in a wrong direction if I'm lazy. This results in a workflow which feels quite a lot different from Claude Code.
how much of that is Opus injecting prior conversations from memory?
I use Claude at work and Kimi for side projects. My org has LiteLLM and Kimi 2.5 enabled but it rarely works, so Claude and GPT are my main tools. I actually enjoy Kimi more as it feels like a dev in a job interview. Watching it reason through problems is a lot like I tend to explain things during whiteboarding sessions. The number of times it says, "wait", is just funny. Claude on the other hand is much more like an employee (or team of employees) that already know they have the job. It doesn't do a ton of explanation up front. (you can dig into processes if you want). It just goes along, asking questions only when it needs... and then delivers a comprehensive report or plan. OpenCode is a better harness. I don't have a direct comparison on costs, as I haven't tried to do the exact same prompt on both models. I can say that I recently had Kimi generate a wrapper around libpq for the ZenC programming language: https://github.com/nobleach/zenc-postgres and it took about an hour or so and cost around 4 dollars.
I am extremely happy with ohmypi, but you could use OpenCode or just keep using Claude Code!
DeepSeek-V4-Pro is adequate plus use DS4-Flash for tasks or other small activity you’d use Haiku or Sonnet for. Go sign up with $10 prepaid.
OpenCode Go - go sign up with $5 for a month and use Qwen-3.7-Max for design/plan/architecture or difficult troubleshooting. Feels closer to Opus 3.6 or 3.7 than DeepSeek, closest I’ve found.
OpenAI Codex, $20 a month plan, use GPT-5.5 via API for the same design/plan/architecture/troubleshooting/author commits. (You can also pay $100 and cut and paste really difficult problems into chat with the GPT-5.5-Pro model.)
Xiaomi MiMo-2.5-Pro, find a friend to give you a $2 referral code, you get 72 cents free. Same pricing as DeepSeek. Somewhere between Sonnet and Opus, quite capable. Apply for the UltraSpeed beta too.
You can switch in and out from these models on the fly in OpenCode or ohmypi and simply find the one that feels best to you. I use CodexBar to watch consumption in near real time.
For a casual user or someone new to programming, Cursor’s $20 plan is an excellent start with Composer-2.5 and Composer-2.5-Fast. You get an API allowance too you can use to access Opus-4.x or GPT-5.5-Pro from OpenCode or ohmypi in addition to Cursor itself.
Finally, if you use Grok or Twitter, SuperGrok at $30 a month has a good vision model, which I used for automated testing of front ends. I’m migrating to locally-run Qwen-3-VL on a commodity Mac, though. If you’re less technical unreach makes hosting local models on a Mac easy.
If you have a powerful GPU like an RTX 5090, try Qwen-3.6 locally on that too. Use ollama or llama-swap which is fairly easy to use.
I have not tried new Kimi yet but we have been able to keep our costs at or below $200 a month per employee with a team of 3 professional developers, 1 graphic designer who uses a lot of Midjourney and Grok Imagine now driven from workflows she made herself in ohmypi, and 1 nontechnical user (account manager / project manager) who uses ohmypi to help her gather requirements and track implementation of them. With a tiny bit of effort we could get that number closer to $75 per employee per month.
Deepseek-V4-Flash-Free on Opencode is what I use most of the time, for simple tasks. Such a good model to give for free (assuming you're okay with harvesting your data)
This was my experience using GLM 5.1 in Claude Code but it works far better in OpenCode, I’d really like to understand why. I think it’s a bit stronger than Sonnet 4.6.
I use the oh-my-openagent planning system and haven’t used vanilla OpenCode enough to know how much that is contributing.
I'm using Claude code + (a patched) litellm proxy + openrouter + Qwen 3.7 max/kimi k2.6/deepseek v4 pro. The only feature that doesn't work is webfetch and web search, which I've replaced with the ddg MCP and a web fetch/search pre hook to redirect the agent. Memory, caching, and everything else works fine.
Qwen comes close to opus for planning but fable is clearly superior. Results for kimi and deepseek are pretty much indistinguishable from opus for coding if opus writes the plan. The biggest difference is output cadence. Kimi for example thinks for a long time then quickly outputs a lot of text.
I'm now testing out fable for research and planning and deepseek v4 flash for coding. I'm guessing results will be pretty similar to opus + deepseek v4 pro and costs should be lower overall.
For some reason I never had a good experience with Kimi (via OpenRouter) in OpenCode. It would only take a few turns for it to run off and mess something up. Terrible instruction following I’d say.
I use DeepSeek V4 Pro now, which works pretty well.
I think there is some threshold after which "best" model doesn't matter, we are not that far from it. Fable now is really good, in a year or so, if Kimi catches up, even if Fable6 is much better, I think I will use kimi at 1/10th of the price.
I said that about opus 4.5 at the time, thinking "this is so good, in 6-12 months the Chinese models will be as good and cheap, I will use them", but I was wrong.. I pay premium for opus4.7/8 and Fable.
But at some point, it will just do the thing you want it to do, and then the race to the bottom will start.
Now that Chinese companies have access to some very good Fable tokens, I hope it speeds up the race.
Exactly, for long running vibe coded stuff that I don't care about quality getting big and smart model is the only option. But for high quality changes where I need to have control and understand everything, where I do everything in small chunks - I can use basic model like Sonnet.
I think the next frontier for competition is speed. Instead of constantly context-switching between multiple agents that I have working on various tasks, I want a single agent that can rip through any prompt in a few seconds, so I can stay in flow on a single task.
yes, that is my point, but at some point, better is unmeasurable, and both the better and the not-as-good produce similar result, and then you pick the one with 1/10th of the price
I was wondering how does Anthropic and likes keep competitive when Opus is ($5 / $25) 5x times more expensive compared to Kimi K2.6 ($0.7 / $3.4) or other Chinese models, while being only marginally better.
My theory is that US enterprise just can't send data to Chinese and that's understandable, but is that "the moat"?
The moat right now is model performance and what that means for how many tokens and additional time you spend.
I say this as a relatively frequent user of Kimi models and generally a big fan. But on not-yet-gamed benchmarks like DeepSWE, Kimi K2.6 is beaten soundly by Claude Sonnet 4.6 ($3 / $15) and even slightly by GPT 5.4 Mini ($0.75 / $4.50).
There's no question Kimi models are very good for a lot of code tasks. They're the best quality open weight model. But to get similar overall outcomes as on Sonnet/Opus, on average you'll spend many more tokens and will have to do more managing of the model. You shouldn't look at price per token, you should look at how much you pay for the entire process.
I'm more interested in how much effort I have to put in, at least while I'm paying in the range of current subscriptions (so ~€100-€200 a month or so). If the prices go up much more than that I'll have to switch to caring more about token efficiency. But at current pricing the bottleneck is my attention, not model efficiency. As such, even a small improvement in model quality - and hence, a decrease in how much attention I have to spend on it - makes a big difference.
I personally dont put any weight to DeepSWE. Other than 5.5 being directionally the best model, it gets the others pretty wrong in my experience. FrontierCode from cognition looks interesting
I think the perception is that it is not 'only marginally better'; whether or not you specifically agree that perceived quality gap lets them differentiate on price.
I'd further say that there are probably enough rational actors running evals out there that the marginally better is not pure vibes for the cases where people are spending lots of money, but I only have direct line of sight to some of those eval suites. Maybe everyone is irrational and anthropic is exploiting that!
I think most people who've tried them both would tell you Anthropic's models are more than marginally better than Kimi. Kimi and the other open source models may score well on SWE-bench or whatever but the gap is noticeable IMHO once you actually try to use them.
It depends on what your task is and how precise your prompts are. Planning with fable or 4.8 and laying out the plan in step by step process and coding with mimo v2.5 pro or dsv4pro or qwen 3.7 max and doing a final review with 5.5 has worked really well for me for infra stuff.
API token price is one thing, but subscriptions on Claude are a good value. Weirdly everyone says that Claude subscriptions are subsidized because of the API price, even though (1) no one actually knows Claude's cost of inference, and (2) Chinese providers are also able to provide cheap inference, so why do they think Claude can't?
I also wonder if Enterprises have deals for other API pricing that is not posted publicly, so all we see is a high API sticker price.
> no one actually knows Claude's cost of inference
There were some rumors stating that their margin is around 70%. So they could go much cheaper probably, talking inference only. The other thing is R&D cost...
I want Opus to be only marginally better, but I do mostly research engineering and its ability to not fuck up my projects is absent. Every time my credits lapse I let kimi and composer2.5 have some play and it’s basically just an excuse for me to keep playing computer because when the oai/ant credits refresh I always need to spend hours recovering from the other models either misconceptions or boneheaded eng practices. Even when I only let it touch my web games…
I reckon right now the Enterprise concern is more FOMO around the AI wave and how to retrain or replace up to hundreds of thousands of employees. I don't think cost is the main concern right now.
But if AI doesn't lead quickly to vast large scale replacement of workers as promised, I could definitely see the C-suits and their gaggle of consultants starting to ask questions about token pricing.
It's only marginally better in the things it's actually comparable to. A\ models are MUCH better in many more things; eg: things Kimi/etc. didn't distill.
Part of Anthropic's moat is Claude Cowork & Claude Code. They got coders comfortable with CC and enterprise users comfortable with Cowork, and both are creating stickiness.
The reality is that $20/$100/$200/mo feels reasonable to a lot of people relative to the value they're getting out of Claude, and if they switch to something else, there's a risk that it won't be as good, and they'll have a new tool to learn.
It's not an insurmountable moat, but don't underestimate the user experience. The iPod didn't win because it was the cheapest device or the one with the most features.
I think none of them having a defacto and high quality English focused cli is a big part of it. None of the Chinese models I've tried have worked well in opensource cli's. Granted, I've only tried a few, but still...
I've been using charm's Crush with GLM for several months and it's been working great. I've only seen it shift to non-english once and it was already in a wonky state when it flipped.
I think any new model not demonstrably maybe 20-30% over Deepseek v4 capabilities priced over the price per token of Deepseek is almost automatically deprecated as low use model (maybe for Planning).
These things enormously benefit from economies of scale. I am fairly certain their margins might be low but they don't actually sell API at loss, however that doesn't mean your cost footprint would be anywhere as low.
I am still very new to the open-weight/source models. If anyone is using them full-time, I’d really love to hear about the setup and how they perform, as I am considering moving my org off Anthropic products.
For personal stuff I use forgecode with openrouter. Firstly, forgecode is a much better harness than Cloude code (IMHO).
Anyway, regarding the models, my experience is that there is not much difference in terms of quality, but the cost difference is insane. At least for how I use agents. Yesterday's example is the following: I am developing a small DSL for search across complex technical documents. I wanted to add a small operator to it and thought that to give fable a spin. It burned through 13 USD and while it delivered the solution it wasn't objectively better than what Deepseek v4 did for 1.7 dollars (same exact task because I was curious).
For full disclosure, I ask agents for piecemeal stuff. Like in the DSL case, I designed the operators and then asked agents to implement them one by one. Probably if I asked to design the whole thing starting from these complex documents Fable would shine, but every time I try to give agents broader scope tasks they burn through millions of tokens, generate questionable code, which I have to spend time familiarize myself with.
you can see that I attribute the models used. What I found was 4.7 was not very good at `go` code which was why you started to see `Gemini 3 Flash` in the attributions.
4.7 is what Cerebras provide and for me, speed in iterations is a lot more important. Having played around with MiMo v2.5.0-Pro, I am 100% sure it could have done what Gemini 3 Flash did.
There were a few points where I was stuck and needed Sonnet to explain things to me, but I think the dirty secret that Anthropic and OpenAI won't tell you is, if you know how to code, the models are honestly good enough.
Based on my experience with MiMo and what others are saying about GLM 5.1, we are now in a hardware race. The Chinese Models are 100% drop in replacement for Claude if you know how to program but want to AI to help amplify what you know. What I will consider now is what provider can provide the fastest inference.
MiMo-v2.5.0-Pro-Ultraspeed is really good at generating good results quickly and burning your money as fast.
I keep trying to switch to the Chinese models, but I keep finding myself asking Claude to fix their outputs. (Both functionality and style.) So I always end up switching back.[0]
I also keep trying GPT, which is quite solid. Very fast, great at debugging. But its code is often overly clever and hurts my brain.
(Maybe fixable with prompting. I tried and it helped the Chinese ones a bit. Just tell them do be elegant, like in the old image AI days "+good -bad"!)
For now I do still need my human brain to actually be able to make sense of the stuff, and Claude is the only one that consistently meets that requirement.
But I am hoping that one of these days, one of the Chinese labs figures out the special sauce :)
--
[0] (For smallish edits, though, I am having a great time with DeepSeek Flash. Practically unlimited AI on tap! How cool is that.)
I have been using deepseek v4 flash as my main model for everything ever since dwarf star came out. I run it on my M4 Max MacBook Pro with 128gb of memory. I run it usually as a server and connect to it over tailscale with my coding machine and use the Pi coding agent. It’s a big leap over using the Qwen models though it doesn’t have vision - so I still will run those when I use vision. GLM 4.7 flash was my previous go to for coding but I’ve completely switched to deepseek for all non-vision things.
These models have open weights, but at the moment most flagship models are practically accessible only through third-party model providers. The main exception is models in the ~30B parameter range, which can still be run on consumer-grade GPUs. That said, even consumer GPUs have become increasingly expensive and difficult to justify in recent years.
I use glm5.1 plus pi with a few customized skills and am very happy with it. I hadn’t touched my Claude 5x plan for a couple of weeks but opened it back up in Claude code when fable was released and did a few tasks and still was happy to return to glm/pi.
I tested it properly and it seems rather decent improvement atleast it does use less tokens for the same task which is good enough a reason for me to use it over k2.6 if I need an open model
Has anyone taken these open weight models from China and stripped the CCP out of them? I do not mean that snarkily, I mean review them thoroughly using techniques for weight introspection (concept activations) in response to things that one might expect would trigger deceptive/malicious behavior if the CCP had actually tried to implant context-specific behaviors (e.g. the accusation of generating vulnerable code if being used in American government applications, which I don't know if it was ever proven).
Just in case there are those who'd reflexively down vote this post, I'd just like to say that in a time of great national geopolitical rivalries, this kind of question is not unreasonable one to ask. Indeed, its applicable question whichever nation you live in.
> Has anyone taken these open weight models from China and stripped the CCP out of them?
The CCP is not influencing my Rust code quality that much. Though I did notice all my lifetimes are now 'static because nothing is ever allowed to leave the party's ownership, unsafe blocks require approval from a central committee.
Honestly the scariest part is that shared mutable state is forbidden unless the state is doing the sharing.
In OpenRouter, there is an "int4" tag for Moonshot provider of Kimi K2. 7 Code. Isn't that too low, particularly coming from the very developer of the model? Os that a mistake? How is it in their direct API offer?
The model is natively quantized (i.e. it was trained that way in the first place, so this is not a post-training quantization which degrades performance).
Output tokens are almost 5x more expensive than mimov2.5 pro/dsv4pro. I’m curious to see if Kimik2.7 is that much better. Feels like kimi are positioning themselves as the premium open source models
I find that I don't use a ton of output tokens. I'm usually around 95% cached input, 4% input, and 1% output.
For me, the big thing with MiMo-V2.5-Pro and DeepSeek V4-Pro is that cached inputs are practically free. Kimi K2.7 Code is 53x more expensive for cached inputs which is 95% of my costs.
If I use 95M cached input tokens, 4M input tokens, and 1M output tokens, that'd be: $18 for cached input on Kimi K2.7 Code vs $0.34 with MiMo/DS; $3.80 for inputs on Kimi vs $1.74 with MiMo/DS; and $4 for output on Kimi vs $0.87 with MiMo/DS.
Of all the places where I'm accumulating costs by using Kimi, it's the cached inputs. The real savings with MiMo/DS's price cut is the cached inputs.
I wish they wouldn't call these "open source" models. The output weights are open but that's more analogous to a binary. The source would be the training data and techniques that went into producing the binary/weights.
"Open weights" is also a term in wide use and accurately tells us what we're getting.
Could be json or non json. Instead of using tools in API, you ask model to share structured output in text. You parse the string to get the JSON. Gives much more control over things you can do.
This maps to what I'm seeing in practice. The gap between demo and production is consistently underestimated, especially around error handling and edge cases.
Deepseek V4 is far from Opus 4.6 level, it might look like it at first glance, but the general reasoning (especially multi-steps) is frankly far off. It's good enough to build great things don't get me wrong, but there is really something that is different from Anthropic models.
Reading their modified license terms, it cracks me up, because they've basically remade the MIT to be the MIT + the one clause that the BSD used to have, which didn't care about MAU or revenue, if you used it in a product, they asked you to 'advertise' them basically. Honestly, its a reasonable request.
This is the cursor callout.
Don't make us shame you into disclosure
Cursor had a specific licensing agreement that allowed them to brand it how they want.
1 reply →
Wasn't the end of that story that Cursor had a non-disclosure licence, so they had not done anything wrong towards Moonshot?
1 reply →
Ah is that what it is? I don't use Cursor, never saw it as being relevant to me, but would not surprise me.
10 replies →
Shaming others when all AI is trained off scraped content and code huh? Many of those sources either breaking ToS or being illegal, such as Anna’s Archive. Bold move. And Chinese models in particular have been accused of distilling off American models.
Don’t you know there’s no honor among thieves?
1 reply →
> they asked you to 'advertise' them basically.
To be clear, the “advertising” clause just requires you to disclose that you use the thing somewhere in the product, such as credits in an “About” section.
I all it advertising clause, because I remember still in the 2000s seeing an Apple ad which at the end of it showed "Unix" or something like that on it, and I remembered that was one of the BSD license requirements, or maybe Apple just did it also just to proudly boast using Unix.
3 replies →
It seems tacked on pretty quickly - I would have expected they try a little more legalese regarding what counts as a "user interface".
Personally, when I use open code or routers, I feel that beyond a certain level, the models don't make a huge difference to me. Except for expensive and mediocre models like Gemini. In that sense, Chinese models are pretty good. I usually write code in function or method units and then design and assemble them together.
GPT series models are more thorough and better, but I'm not sure if the difference is enormous. It seems to depend on the workflow, but in my opinion, if you are thorough enough, I wonder if there really is a big difference
I've kind of given up on the routers for "free" inference, as you would expect, they tend to give you sub-par thinking because they are obviously trying to conserve as much inference as possible.
I've had some success turning my macbook M1 pro into a heating pad with Qwen 3.6 35B A3B MTP. Trying to use Gemini models "locally" resulted in a similar "short shrift" of effort resulting in mistakes and lots of turns. The reports of Fable being relentlessly "proactive" shows you can go the other direction as well, if you have strong enough branding and effective invoicing.
> I've kind of given up on the routers for "free" inference, as you would expect, they tend to give you sub-par thinking because they are obviously trying to conserve as much inference as possible.
Xiaomi MiMo ($6/mo: https://platform.xiaomimimo.com/token-plan) & Alibaba Qwen ($50/mo: https://www.alibabacloud.com/en/campaign/ai-scene-coding) have generous limits on fixed subscriptions.
5 replies →
> The reports of Fable being relentlessly "proactive"
For the curious: https://news.ycombinator.com/item?id=48498573 - “Claude Fable is relentlessly proactive”.
Tangent: did the MTP help you at all? I’ve tested that model back to back on my M1 Max MBP and the MTP version was actually marginally worse. I wonder if I didn’t use the right settings, although I tried several based on the obvious sources.
In my experience, there's little difference between implementing individual functions between frontier models and SotA ~30B param models.
Once you have a coherent design (the hard part), you can feed it to a pretty small model and get basically the same quality.
They'll not one-shot, but they're faster and cheaper, so it still works out in your favor.
Plus you can do it locally...
I have a similar experience. However, when including code review, I think the GPT model is the most impressive
The difference in outcome isn't that big but yes, you need to be more rigorous. For instance I've found that the Kimi K2.5 and K2.6 models will comment out failing tests rather than fix a problem they just caused (mistaking them for "pre-existing failures"), so you need to specifically make commented-out tests break the build. I've not personally had that problem with any of the Anthropic or OpenAI models.
I wonder why it's the natural tendency of models to BS or do stuff like this when they don't have the correct answer - it's clear that they can program refusal into them, but for some reason, refusal has to be injected after the fact, and models can't really arrive at the conclusion that they can't answer properly.
2 replies →
I really hope we stop using the term "Chinese models". It has this air of Negative connotation. It's the equivalent of calling cars Japanese, which people used to do but now is almost entirely meaningless. You just call them Toyota, Honda, Lexus etc.
I don't think "Chinese" is pejorative in this context any more than "American" is. They are one of the two ecosystems. What's wrong with saying "Japanese cars" today?
9 replies →
For me, it has a positive connotation! In my experience, Chinese Model means cheaper, but still quite effective model you can use for millions of tokens without burning your entire wallet in seconds. That's why I get more excited over a Chinese model release over American models.
Japanese cars is actually a positive qualifier. I'd say anything Japanese motor-powered.
1 reply →
No thanks.
The term seems to have the connotation of "competitive at 1/10 the price of Claude", so I don't see the problem.
It's not Harbor Freight Chinese (and heck even they have decent stuff sometimes now too).
You don't think people still talk about Japanese cars as a distinction in quality from US or European ones?
I don't know, I tried using one of the Chinese models and it was VERY quick to scan my entire home dir, so maybe your threat surface is a little different than mine
1 reply →
[flagged]
45 replies →
[flagged]
You are right. I agree.It may seem like a kind of bias, but I hadn't thought of that part. Thank you for pointing out my bias.
2 replies →
I just had Kimi K2.7-code rebase my Fil-C OpenSSL patch from 3.3.1 to 3.5.7 with quite bare bones instructions and it seems to have worked.
177KB patch, so it's not a small change. The patch did not apply cleanly initially; the agent had to do nontrivial work.
I just showed it the patch against 3.3.1, what command to use to build, and the path to 3.5.7 along with a link to the documentation of the change (https://fil-c.org/constant_time_crypto).
Note, I use my own coding agent (T800, which isn't public, and was previously well tested and tuned for K2.5).
I think this cost me between $5 and $10 in API usage.
(EDIT: OpenSSL, not OpenSSH)
"T800"
Do you have your agent say things like "Hasta la vista baby", or "I'll be back, after I clear my context" ?
Yes
I would really love to know if anyone has any experience with something like opencode + Kimi K2.6/2.7 now compared to Claude Code. What is better, what is worse, what is the cost comparison. I am currently paying $100 for the 5x Max plan, but Fable is running through the usage limits quite drastically and I cannot really say it's night and day compared to Opus. Also, I use this mostly for my side projects, so the $100 bill is quite noticeable. I definitely don't want to pay more.
I do have this experience. I've used Claude Code (with Opus mostly), and then switched to opencode (mostly with Kimi 2.6) for my personal projects; it's based on a couple months of use.
Claude Code is better. But Opencode + kimi 2.6 is workable, which is big. For bare code writing, if you know what exactly you want, most popular models are fine (deepseek, kimi, etc), it feels more or less the same as anthropic models.
At the same time, Opus seems to understand my intent way better than e.g. deepseek. I need to be much more precise with my prompts when using deepseek - it often goes in a wrong direction if I'm lazy. This results in a workflow which feels quite a lot different from Claude Code.
Kimi is in between - for me it brings back "lazy prompting" workflow, and I can trust its plans more than deepseek. It enables a workflow similar to Claude Code, it's workable, but it is a bit worse everywhere. Smaller context, a bit more errors, decisions are a bit worse, recommendations are a bit worse, debugging capabilities are a bit worse, etc.
On the usage side, $100 Claude plan is a great value actually. On paper, per-token kimi is way cheaper, but Claude subscriptions are heavily subsidized - you get much more tokens than $100 can buy you. So, in the end, opencode + kimi vs claude code could be of a similar cost, for similar usage patterns. Deepseek can be cheaper, and it has insanely cheap cached tokens, but experience may vary - depending on your habits, you may need to adjust how you work, coming from claude code.
I'd say for side projects something like $10 Opencode Go plan + $10 of extra DeepSeek v4 credits (e.g. on OpenRouter) can be very workable.
To my experience claude/codex $20 are even more subsidized, so running on sonnet or gpt5.4 again gives you more usage.
4 replies →
This is generally been my experience as well, but i think the main reason for claude code being better at understanding intent is their massive system prompt.
>At the same time, Opus seems to understand my intent way better than e.g. deepseek. I need to be much more precise with my prompts when using deepseek - it often goes in a wrong direction if I'm lazy. This results in a workflow which feels quite a lot different from Claude Code.
how much of that is Opus injecting prior conversations from memory?
2 replies →
according to this opencode and cursor cli perform better than claude code: https://x.com/kunchenguid/status/2065345999682568593
1 reply →
I use Claude at work and Kimi for side projects. My org has LiteLLM and Kimi 2.5 enabled but it rarely works, so Claude and GPT are my main tools. I actually enjoy Kimi more as it feels like a dev in a job interview. Watching it reason through problems is a lot like I tend to explain things during whiteboarding sessions. The number of times it says, "wait", is just funny. Claude on the other hand is much more like an employee (or team of employees) that already know they have the job. It doesn't do a ton of explanation up front. (you can dig into processes if you want). It just goes along, asking questions only when it needs... and then delivers a comprehensive report or plan. OpenCode is a better harness. I don't have a direct comparison on costs, as I haven't tried to do the exact same prompt on both models. I can say that I recently had Kimi generate a wrapper around libpq for the ZenC programming language: https://github.com/nobleach/zenc-postgres and it took about an hour or so and cost around 4 dollars.
I am extremely happy with ohmypi, but you could use OpenCode or just keep using Claude Code!
DeepSeek-V4-Pro is adequate plus use DS4-Flash for tasks or other small activity you’d use Haiku or Sonnet for. Go sign up with $10 prepaid.
OpenCode Go - go sign up with $5 for a month and use Qwen-3.7-Max for design/plan/architecture or difficult troubleshooting. Feels closer to Opus 3.6 or 3.7 than DeepSeek, closest I’ve found.
OpenAI Codex, $20 a month plan, use GPT-5.5 via API for the same design/plan/architecture/troubleshooting/author commits. (You can also pay $100 and cut and paste really difficult problems into chat with the GPT-5.5-Pro model.)
Xiaomi MiMo-2.5-Pro, find a friend to give you a $2 referral code, you get 72 cents free. Same pricing as DeepSeek. Somewhere between Sonnet and Opus, quite capable. Apply for the UltraSpeed beta too.
You can switch in and out from these models on the fly in OpenCode or ohmypi and simply find the one that feels best to you. I use CodexBar to watch consumption in near real time.
For a casual user or someone new to programming, Cursor’s $20 plan is an excellent start with Composer-2.5 and Composer-2.5-Fast. You get an API allowance too you can use to access Opus-4.x or GPT-5.5-Pro from OpenCode or ohmypi in addition to Cursor itself.
Finally, if you use Grok or Twitter, SuperGrok at $30 a month has a good vision model, which I used for automated testing of front ends. I’m migrating to locally-run Qwen-3-VL on a commodity Mac, though. If you’re less technical unreach makes hosting local models on a Mac easy.
If you have a powerful GPU like an RTX 5090, try Qwen-3.6 locally on that too. Use ollama or llama-swap which is fairly easy to use.
I have not tried new Kimi yet but we have been able to keep our costs at or below $200 a month per employee with a team of 3 professional developers, 1 graphic designer who uses a lot of Midjourney and Grok Imagine now driven from workflows she made herself in ohmypi, and 1 nontechnical user (account manager / project manager) who uses ohmypi to help her gather requirements and track implementation of them. With a tiny bit of effort we could get that number closer to $75 per employee per month.
Deepseek-V4-Flash-Free on Opencode is what I use most of the time, for simple tasks. Such a good model to give for free (assuming you're okay with harvesting your data)
> I am extremely happy with ohmypi, but you could use OpenCode or just keep using Claude Code!
What's the benefit of using OMP over OpenCode?
Just the sheer amount of options in OMP overwhelmed me. But I also use both via ACP in Zed so the CLI itself doesn't matter much.
4 replies →
Also, if you do have SuperGrok, forget using Grok, they are giving you Composer 2.5 in Grok Build.
I just switched from Llama.cpp to Llama swap with the help of codex. It was great.
I need to try the DSv4 stuff sometime.
I can only talk about GLM 5.1 which is roughly at sonnet 4 levels imo.
It's good, does most tasks well that I throw at it, but will fail at anything congitive/complex. It gets stuck often. It costs ~6$ a month though
This was my experience using GLM 5.1 in Claude Code but it works far better in OpenCode, I’d really like to understand why. I think it’s a bit stronger than Sonnet 4.6.
I use the oh-my-openagent planning system and haven’t used vanilla OpenCode enough to know how much that is contributing.
1 reply →
I'm using Claude code + (a patched) litellm proxy + openrouter + Qwen 3.7 max/kimi k2.6/deepseek v4 pro. The only feature that doesn't work is webfetch and web search, which I've replaced with the ddg MCP and a web fetch/search pre hook to redirect the agent. Memory, caching, and everything else works fine.
Qwen comes close to opus for planning but fable is clearly superior. Results for kimi and deepseek are pretty much indistinguishable from opus for coding if opus writes the plan. The biggest difference is output cadence. Kimi for example thinks for a long time then quickly outputs a lot of text.
I'm now testing out fable for research and planning and deepseek v4 flash for coding. I'm guessing results will be pretty similar to opus + deepseek v4 pro and costs should be lower overall.
The Kimi problem is it doesn’t follow instructions and goes off track often.
Other than that it’s pretty decent (for the price).
Sounds like it was distilled from Claude. I don't understand the appeal of an agent that does whatever it wants.
2 replies →
This. It will try to fix and refactor things that don’t need fixing because it gets stuck trying to solve the problem at hand.
Yup. I’m hoping this variant fixes these issues.
For some reason I never had a good experience with Kimi (via OpenRouter) in OpenCode. It would only take a few turns for it to run off and mess something up. Terrible instruction following I’d say.
I use DeepSeek V4 Pro now, which works pretty well.
The best is GLM (though it's not as cheap as DeepSeek or Kimi) and use it with Claude Code.
I think there is some threshold after which "best" model doesn't matter, we are not that far from it. Fable now is really good, in a year or so, if Kimi catches up, even if Fable6 is much better, I think I will use kimi at 1/10th of the price.
I said that about opus 4.5 at the time, thinking "this is so good, in 6-12 months the Chinese models will be as good and cheap, I will use them", but I was wrong.. I pay premium for opus4.7/8 and Fable.
But at some point, it will just do the thing you want it to do, and then the race to the bottom will start.
Now that Chinese companies have access to some very good Fable tokens, I hope it speeds up the race.
Depending on who you are and how you use these models, we're already at this point
Exactly, for long running vibe coded stuff that I don't care about quality getting big and smart model is the only option. But for high quality changes where I need to have control and understand everything, where I do everything in small chunks - I can use basic model like Sonnet.
1 reply →
I think the next frontier for competition is speed. Instead of constantly context-switching between multiple agents that I have working on various tasks, I want a single agent that can rip through any prompt in a few seconds, so I can stay in flow on a single task.
price/token isnt the only thing relevant. if you have to ask the AI again, it'll cost you more than when it gets things right in the first place.
so better models may still be cheaper even if the price per token is higher.
yes, that is my point, but at some point, better is unmeasurable, and both the better and the not-as-good produce similar result, and then you pick the one with 1/10th of the price
I was wondering how does Anthropic and likes keep competitive when Opus is ($5 / $25) 5x times more expensive compared to Kimi K2.6 ($0.7 / $3.4) or other Chinese models, while being only marginally better.
My theory is that US enterprise just can't send data to Chinese and that's understandable, but is that "the moat"?
The moat right now is model performance and what that means for how many tokens and additional time you spend.
I say this as a relatively frequent user of Kimi models and generally a big fan. But on not-yet-gamed benchmarks like DeepSWE, Kimi K2.6 is beaten soundly by Claude Sonnet 4.6 ($3 / $15) and even slightly by GPT 5.4 Mini ($0.75 / $4.50).
There's no question Kimi models are very good for a lot of code tasks. They're the best quality open weight model. But to get similar overall outcomes as on Sonnet/Opus, on average you'll spend many more tokens and will have to do more managing of the model. You shouldn't look at price per token, you should look at how much you pay for the entire process.
I'm more interested in how much effort I have to put in, at least while I'm paying in the range of current subscriptions (so ~€100-€200 a month or so). If the prices go up much more than that I'll have to switch to caring more about token efficiency. But at current pricing the bottleneck is my attention, not model efficiency. As such, even a small improvement in model quality - and hence, a decrease in how much attention I have to spend on it - makes a big difference.
I personally dont put any weight to DeepSWE. Other than 5.5 being directionally the best model, it gets the others pretty wrong in my experience. FrontierCode from cognition looks interesting
I'm not sure I would put too much weight on DeepSWE as a benchmark, given that GPT-5.4-mini ended up close to Opus 4.6 there.
3 replies →
I think the perception is that it is not 'only marginally better'; whether or not you specifically agree that perceived quality gap lets them differentiate on price.
I'd further say that there are probably enough rational actors running evals out there that the marginally better is not pure vibes for the cases where people are spending lots of money, but I only have direct line of sight to some of those eval suites. Maybe everyone is irrational and anthropic is exploiting that!
I think most people who've tried them both would tell you Anthropic's models are more than marginally better than Kimi. Kimi and the other open source models may score well on SWE-bench or whatever but the gap is noticeable IMHO once you actually try to use them.
It depends on what your task is and how precise your prompts are. Planning with fable or 4.8 and laying out the plan in step by step process and coding with mimo v2.5 pro or dsv4pro or qwen 3.7 max and doing a final review with 5.5 has worked really well for me for infra stuff.
1 reply →
> My theory is that US enterprise just can't send data to Chinese
Lots of US providers are hosting these “open source” models so doubt that’s the problem.
API token price is one thing, but subscriptions on Claude are a good value. Weirdly everyone says that Claude subscriptions are subsidized because of the API price, even though (1) no one actually knows Claude's cost of inference, and (2) Chinese providers are also able to provide cheap inference, so why do they think Claude can't?
I also wonder if Enterprises have deals for other API pricing that is not posted publicly, so all we see is a high API sticker price.
> no one actually knows Claude's cost of inference
There were some rumors stating that their margin is around 70%. So they could go much cheaper probably, talking inference only. The other thing is R&D cost...
I only have knowledge of one enterprise deal but there is no discount. Which I found surprising.
I want Opus to be only marginally better, but I do mostly research engineering and its ability to not fuck up my projects is absent. Every time my credits lapse I let kimi and composer2.5 have some play and it’s basically just an excuse for me to keep playing computer because when the oai/ant credits refresh I always need to spend hours recovering from the other models either misconceptions or boneheaded eng practices. Even when I only let it touch my web games…
You have to revert to Opus 4.5 and 4.6. I bet you'll see a massive improvement based on what you're describing
Your question relies on the premise that Chinese companies continue releasing free models. What's "the moat" for them continuing to do that?
I reckon right now the Enterprise concern is more FOMO around the AI wave and how to retrain or replace up to hundreds of thousands of employees. I don't think cost is the main concern right now.
But if AI doesn't lead quickly to vast large scale replacement of workers as promised, I could definitely see the C-suits and their gaggle of consultants starting to ask questions about token pricing.
> while being only marginally better.
It's only marginally better in the things it's actually comparable to. A\ models are MUCH better in many more things; eg: things Kimi/etc. didn't distill.
For those things the difference is like a cliff.
That's a baseless claim that borderline reads like shilling. Do you have any proof of that you wrote there?
Part of Anthropic's moat is Claude Cowork & Claude Code. They got coders comfortable with CC and enterprise users comfortable with Cowork, and both are creating stickiness.
The reality is that $20/$100/$200/mo feels reasonable to a lot of people relative to the value they're getting out of Claude, and if they switch to something else, there's a risk that it won't be as good, and they'll have a new tool to learn.
It's not an insurmountable moat, but don't underestimate the user experience. The iPod didn't win because it was the cheapest device or the one with the most features.
Performance. I pay for Opencode but none of the models give me Codex performance, so I have to keep my 20€ subscription+ the Opencode one
I think none of them having a defacto and high quality English focused cli is a big part of it. None of the Chinese models I've tried have worked well in opensource cli's. Granted, I've only tried a few, but still...
I've been using charm's Crush with GLM for several months and it's been working great. I've only seen it shift to non-english once and it was already in a wonky state when it flipped.
i use github copilot cli + openrouter + qwen 3.7 max and it's really much better than i expected (used to opus 4.7 at work)
huh? They all work great in omp/opencode unless you mean their own native clis like kimi code
[dead]
I think any new model not demonstrably maybe 20-30% over Deepseek v4 capabilities priced over the price per token of Deepseek is almost automatically deprecated as low use model (maybe for Planning).
DeepSeek v4 Pro is not actually that good a model compared to GLM 5.1 and Kimi K2.6. It's an okay coder/thinker for the price.
How so? In my experience trying these models using opencode Go, DeepSeek is superior to GLM 5.1.
If anything, DS4 has 1 million context window, while GLM 5.1 has 200K.
There are also benchmarks comparing the two: https://artificialanalysis.ai/models/comparisons/deepseek-v4...
Is Deepseek just eating cost or are people able to host their open models for comparable costs?
If openrouter is to be trusted, the cheapest offers that are not from Deepseek itself are:
- twice as expensive on the output (1.52 vs 0.87)
- six times as expensive on the input (0.33 vs 0.05)
https://openrouter.ai/deepseek/deepseek-v4-pro?sort=price#pr...
Other people are hosting it in the same order of magnitude. Xioami recently matched DeepSeek’s pricing.
They focused on caching and other optimizations.
These things enormously benefit from economies of scale. I am fairly certain their margins might be low but they don't actually sell API at loss, however that doesn't mean your cost footprint would be anywhere as low.
Likely CCP-subsidized
I am still very new to the open-weight/source models. If anyone is using them full-time, I’d really love to hear about the setup and how they perform, as I am considering moving my org off Anthropic products.
Anecdotal, but here's my experience.
For personal stuff I use forgecode with openrouter. Firstly, forgecode is a much better harness than Cloude code (IMHO).
Anyway, regarding the models, my experience is that there is not much difference in terms of quality, but the cost difference is insane. At least for how I use agents. Yesterday's example is the following: I am developing a small DSL for search across complex technical documents. I wanted to add a small operator to it and thought that to give fable a spin. It burned through 13 USD and while it delivered the solution it wasn't objectively better than what Deepseek v4 did for 1.7 dollars (same exact task because I was curious).
For full disclosure, I ask agents for piecemeal stuff. Like in the DSL case, I designed the operators and then asked agents to implement them one by one. Probably if I asked to design the whole thing starting from these complex documents Fable would shine, but every time I try to give agents broader scope tasks they burn through millions of tokens, generate questionable code, which I have to spend time familiarize myself with.
I'm making DSLs a lot as an architecture pattern also. I'd be curious to know what stack you're using this and how you're approaching it
2 replies →
I created this and I would say glm-4.7 accounts for 80% of the code in https://github.com/gitsense/gsc-cli
If you look at a file like:
https://github.com/gitsense/gsc-cli/blob/main/internal/cli/r...
you can see that I attribute the models used. What I found was 4.7 was not very good at `go` code which was why you started to see `Gemini 3 Flash` in the attributions.
4.7 is what Cerebras provide and for me, speed in iterations is a lot more important. Having played around with MiMo v2.5.0-Pro, I am 100% sure it could have done what Gemini 3 Flash did.
There were a few points where I was stuck and needed Sonnet to explain things to me, but I think the dirty secret that Anthropic and OpenAI won't tell you is, if you know how to code, the models are honestly good enough.
Based on my experience with MiMo and what others are saying about GLM 5.1, we are now in a hardware race. The Chinese Models are 100% drop in replacement for Claude if you know how to program but want to AI to help amplify what you know. What I will consider now is what provider can provide the fastest inference.
MiMo-v2.5.0-Pro-Ultraspeed is really good at generating good results quickly and burning your money as fast.
I keep trying to switch to the Chinese models, but I keep finding myself asking Claude to fix their outputs. (Both functionality and style.) So I always end up switching back.[0]
I also keep trying GPT, which is quite solid. Very fast, great at debugging. But its code is often overly clever and hurts my brain.
(Maybe fixable with prompting. I tried and it helped the Chinese ones a bit. Just tell them do be elegant, like in the old image AI days "+good -bad"!)
For now I do still need my human brain to actually be able to make sense of the stuff, and Claude is the only one that consistently meets that requirement.
But I am hoping that one of these days, one of the Chinese labs figures out the special sauce :)
--
[0] (For smallish edits, though, I am having a great time with DeepSeek Flash. Practically unlimited AI on tap! How cool is that.)
I have been using deepseek v4 flash as my main model for everything ever since dwarf star came out. I run it on my M4 Max MacBook Pro with 128gb of memory. I run it usually as a server and connect to it over tailscale with my coding machine and use the Pi coding agent. It’s a big leap over using the Qwen models though it doesn’t have vision - so I still will run those when I use vision. GLM 4.7 flash was my previous go to for coding but I’ve completely switched to deepseek for all non-vision things.
These models have open weights, but at the moment most flagship models are practically accessible only through third-party model providers. The main exception is models in the ~30B parameter range, which can still be run on consumer-grade GPUs. That said, even consumer GPUs have become increasingly expensive and difficult to justify in recent years.
You can definitely go above 30B on consumer hardware – 2x gpus, spark, mac, half byte quants etc.
I use glm5.1 plus pi with a few customized skills and am very happy with it. I hadn’t touched my Claude 5x plan for a couple of weeks but opened it back up in Claude code when fable was released and did a few tasks and still was happy to return to glm/pi.
Better than Qwen3.6-35B-A3B-8bit ?
When I tried glm found it way way slower (omlx as runtime)
1 reply →
Qwen 3.6 seems to be the strongest local models, works OK on an RTX 5090 or a > 32GB Mac.
I used glm5/5.1 for 60 days. Certainly better than Sonnet 4.6, not as good as Opus or GPT.
Use DCP or Magic Context plugin in OpenCode to keep the context below 160k and you're fine.
I tested it properly and it seems rather decent improvement atleast it does use less tokens for the same task which is good enough a reason for me to use it over k2.6 if I need an open model
Has anyone taken these open weight models from China and stripped the CCP out of them? I do not mean that snarkily, I mean review them thoroughly using techniques for weight introspection (concept activations) in response to things that one might expect would trigger deceptive/malicious behavior if the CCP had actually tried to implant context-specific behaviors (e.g. the accusation of generating vulnerable code if being used in American government applications, which I don't know if it was ever proven).
Just in case there are those who'd reflexively down vote this post, I'd just like to say that in a time of great national geopolitical rivalries, this kind of question is not unreasonable one to ask. Indeed, its applicable question whichever nation you live in.
> Has anyone taken these open weight models from China and stripped the CCP out of them?
The CCP is not influencing my Rust code quality that much. Though I did notice all my lifetimes are now 'static because nothing is ever allowed to leave the party's ownership, unsafe blocks require approval from a central committee.
Honestly the scariest part is that shared mutable state is forbidden unless the state is doing the sharing.
Otherwise it is pretty ok.
Check out TNG on huggingface
They are a consultancy in Germany, but I watched a presentation on them tuning and removing bias from Deepseek models. It was quite interesting.
https://www.tngtech.com/en/about-us/news/release-of-deepseek...
(I upvoted your question as I agree)
Its not just code we need to worry about, its also subliminal messaging and other things.
Sounds like something that heretic or similar might be useful for?
https://github.com/p-e-w/heretic
Eh even corporate created LLMs are suspect to corporate biases. Nothing is safe.
Everything is the same is not a serious argument because they are not the same.
3 replies →
In OpenRouter, there is an "int4" tag for Moonshot provider of Kimi K2. 7 Code. Isn't that too low, particularly coming from the very developer of the model? Os that a mistake? How is it in their direct API offer?
The model is natively quantized (i.e. it was trained that way in the first place, so this is not a post-training quantization which degrades performance).
Isn't it not completely quantized? I thought there were some dense parts but most is int4?
1 reply →
But the huggingface link mentions BF16, F16, and I32?
2 replies →
Output tokens are almost 5x more expensive than mimov2.5 pro/dsv4pro. I’m curious to see if Kimik2.7 is that much better. Feels like kimi are positioning themselves as the premium open source models
I find that I don't use a ton of output tokens. I'm usually around 95% cached input, 4% input, and 1% output.
For me, the big thing with MiMo-V2.5-Pro and DeepSeek V4-Pro is that cached inputs are practically free. Kimi K2.7 Code is 53x more expensive for cached inputs which is 95% of my costs.
If I use 95M cached input tokens, 4M input tokens, and 1M output tokens, that'd be: $18 for cached input on Kimi K2.7 Code vs $0.34 with MiMo/DS; $3.80 for inputs on Kimi vs $1.74 with MiMo/DS; and $4 for output on Kimi vs $0.87 with MiMo/DS.
Of all the places where I'm accumulating costs by using Kimi, it's the cached inputs. The real savings with MiMo/DS's price cut is the cached inputs.
95/4/1 holds here too
It's not more expensive at all. They are all open weights models. I run them on 2x8xH100. They cost the same.
Openrouter has them as significantly more expensive.
Benchmark geometric mean
- GPT-5.5: 62.7%
- Opus 4.8: 62.2%
- Kimi K2.7 Code: 56.3%
- Kimi K2.6: 48.2%
Would be nice to have 5.2 and 4.6 for comparison.
Seems to be similar level to Kimi K.26, just that it's more token efficient and cheaper to run:
https://aibenchy.com/compare/moonshotai-kimi-k2-6-medium/moo...
I wish they wouldn't call these "open source" models. The output weights are open but that's more analogous to a binary. The source would be the training data and techniques that went into producing the binary/weights.
"Open weights" is also a term in wide use and accurately tells us what we're getting.
It's not quite as closed as a binary, it is very standard practice to take these models and fine-tune them.
If there were actually even close to frontier open source models, this would be more of a discussion, but everyone knows these mean open weight.
Is this Moonshot.ai's attempt to replicate Composer 2.5 (coding fine-tune of Kimi 2.5) from Cursor IDE?
Great! Finally follows custom tool call format (k2.6 couldn't). It's a good indicator of instructions following and agentic behaviour.
UIs it's generating is pretty good, not without problems, but certainly better than other models at this price point.
What do you mean by custom format? Non-json?
Could be json or non json. Instead of using tools in API, you ask model to share structured output in text. You parse the string to get the JSON. Gives much more control over things you can do.
For example model shares
<tool_call name="getWeather"> <param name="city">London</param> </tool_call>
This maps to what I'm seeing in practice. The gap between demo and production is consistently underestimated, especially around error handling and edge cases.
I think deepseek has crossed the threshold for being on par with opus 4.6 and kimi is doing a great job in shipping velocity.
Deepseek V4 is far from Opus 4.6 level, it might look like it at first glance, but the general reasoning (especially multi-steps) is frankly far off. It's good enough to build great things don't get me wrong, but there is really something that is different from Anthropic models.
agreed
How is 2.7 a thing _now_ ? it's not even mentioned on moonshot's webpage..
It's not 2.7. It's 2.7-Code, and it's 2.6 token-optimised for coding.
https://platform.kimi.ai/docs/guide/kimi-k2-7-code-quickstar...
Looks interesting but yet no Ollama model?
insanely great!
[flagged]
[flagged]
[flagged]