Comment by Xeoncross
3 days ago
If you have a 24-64GB mac, consider running Qwen3.8 27B locally at night. It's a bit slower to run locally, but if you're sleeping it's less of a problem.
Depending on your memory, you'll need to use the weaker Q4 versions but they still perform well.
It ranks higher than GPT-5.3 Codex (xhigh) or Claude Opus 4.6 (max) so is great for pairing with https://github.com/kunchenguid/gnhf for nightly experimentation, cleanup, or recommendation lists for in the morning.
I have all sorts of local compute, and local models fairly capable the Frontier models still way more capable/faster and local electricity consumption is something else. Good thing it is getting colder around here.
I often pair them up, and I have an Astra or Sol work as a supervisor and reviewer while Qwen 27B FP8 or Qwen 3.8 Flash Next implements things. I mostly do it as an experiment, just to see what kind of level of autonomy I can get, and they are slow to getting a decent reviewed outcome despite Qwen27B running at 100+ tps and 3.5-4K prefill rates and Qwen3.8 Next at 40 tps and 1-2K prefill. I've been also using similar approach more with OMP, not just the straight Pi harness. And OMP seems to be slower because it has more guardrails. OMP has an interesting feature where you can assign a better LLM as an advisor, wehere it just sort of monitors the progress and injects guidance. And it definitely helps, but one has to be careful. It actually turns out to be expensive if the cache reads are expensive. I learned it the hard way. Where on Fireworks' API, the cache rates for GLM 5.3 flash are quite a bit more expensive than for DeepSeek, and a simple runs ended up costing me three bucks in oversight. So a better way is to have a Frontier model running a separate tmux pane and just directing it to Wake up every 10 minutes, take a peek at what's going on, review the milestones give feedback and then sleep. This turns out to be pretty decent cost saving strategy when quote needs to be stretched. Paradoxically, OpenAI tightening up their quota allowance once they released Astra actually pushed me into all these sorts of experiments, and it's actually been interesting. I've been exploring all these smaller flash models, and it's been nice. I do like using local LLMs for chore type tasks that are just mostly information gathering, post-session reviews, stuff like that.
There also exists a $600-800 GPU that can run Qwen 27B 3.8 @ like 60 t/s for around 200W of energy.
I find Qwen Flash Next quite competent as well. 27B is a solid worker like you said. If you batch work and let them crank they do remarkably well. I am working on.
I started using Herdr and taught my agents to use it. So I use OMP loop and or goal, and it has a review cycles to wake up an Opus or Sol reviewer to make sure nothing is going off the rails with a local qwen flash next coordinating for me. I kinda prefer Sol, it seems like a more patient and thorough model, especially Sol 6, but Opus 5.5 is really good and its voice and attitude is not as grating as Opus 5 for sure.
I have a few V100 GPU running Qwen 27B and they do all the work overnight. Not quite the same speeds you have yet, but this is V100 machine and an old gaming machine with 16GB 4080 and a handful of 32GB V100s... all in less than 3K (ignoring that my gaming machine is 3 years old, but runs qwen flash next for free now as I game not a lot) for my little "we have AI at home" projects and there is a lot of interest in these old GPU now because they are rolling out of data centers now.
For my local work and personal projects... they just seem to be getting done in this setup. Every few days I sit down and do a big cycle with astra/fable/opus batch things up. I have projects that are basically "i want to see what happens" to "I want this to be good, I understand the code". Some of the throwaway projects that have just kind of magically finished more or less how I wanted have been great.
What do you mean by "local electricity consumption is something else"? Doesn't an M3 Pro for example draw about as much power as a bright incandescent lightbulb for a maxed out gpu workload (~100W)? That's less than a tenth of what a frontier model will use in the datacenter (which I believe are racks of BlackWell or Vera Lynn GPUs, each using 500W+).
Because M3 Pro is not a real thing as far as actual agentic workflows go. I run dual R9700 boxes those idle at 150W Because of a Ryzen AM5, and I run Spark boxes, which are decent But still idle at 45-50 watts a pop. My favorite, to be honest, is a 5090 with a Qwen 27B because that one is good for quick hitters and flies, But again, the box itself idles at 140 watts. So it's really the idle power that I don't like And it's too much of an inconvenience to power boxes down and power them back on, so they just end up running and sucking electricity. I am working on getting some sort of a smarter proxy setup where I would give boxes time to wake up and go to the cloud while they're waking up. So the whole point of having these things running in the background is that you do want them to be almost always available. So power consumption is definitely an issue.
On my 64gb m3 max qwen3.8 27b has been great for planning and then letting qwen3.6 35ba3b actually implement the planned changes.
You might be interested in https://prismml.com/news/bonsai-2-27b
This looked exciting until I read that it gets stuck in loops and generally wasn't a useful model
Please cite a reputable source for this claim, and keep in mind we are talking about Bonsai 2 and not the original Bonsai.
Even on the original one, the authors have replied to such issues with working solutions: https://github.com/PrismML-Eng/Bonsai-demo/issues/131
for 32gb, this is my model of reference now, you have to run it with a patched version of llama and is still not available in lmstudio or omlx, waiting for that to streamline the experience a bit. But so far, the best i had till now.
It looks like upstream work is underway: https://github.com/ggml-org/llama.cpp/issues/29058
The sheer number of geniuses working so fast on this repo never ceases to amaze me.
If you're on MacOS, with atleast an M3 chip and 32GB, you should look at the splash engine.
GNHF seems exactly what I've been aiming for to handle overnight tasks.
It's _so_ good, I no longer bother with Sonnet and use it locally for everything.
Consider bumping reasoning down to Medium as a default though, I agree with simonw it over thinks https://simonwillison.net/2026/Aug/16/qwen-38-27b/
Isn't there a way to set up a thinking budget so it automatically tells the model "Conclude your reasoning now and provide the final answer." ?
I still haven't found a use case for Qwen3.8 27B that Qwen 3.6 35b A3b (MoE) is better at. I can get at most 40 tokens/second with 27B, but I get around 90 tokens/second with the MoE and it seems to be a more capable model.
I guess I should still keep experimenting though. Maybe I'm just not using a dense model correctly.
What would you personally recommend for those that have 128 GB?
I have a Ryzen AI Max+ 395 with 128 GB for running those sort of "background task", and my sweet spot is currently Qwen3.8 Flash-Next IQ4 at 96 GB.
if you have 128 GB, you could use Qwen Flash Next at some reasonable quant, with the new SSD hack for only keeping some of the model resident in memory.