Comment by SamInTheShell
10 hours ago
From what I’ve been seeing, the Mac studios do look like they have potential. I was looking to drop $10k-$15k on one until recently. After comparing a Radeon 7900 XTX vs Ryzen Halos 128GB vs M1 MacBook Pro 64Gb, I landed on just getting an external closure setup with Nvidia RTX 5090.
The model I’m specifically targeting to use at high speeds is Qwen 3.8 27b @q4ks. This model actually proved to be good at coding (it sits somewhere between Sonnet 5 and Opus 5 capability). M1 got 10 tok/s, Ryzen Halo 20tok/s, and Radeon 7900 XTX 50tok/s (can only do 128k context window in Radeon card).
The prefill gets extremely slow around 50k tokens in context window (whatever prompt processing stage entails could be wrong about phases here). It takes about 2 hours to fill the context.
Even with a drafter model intended for speed instead of mtp, I can’t get past 70tok/s, still is extremely slow to process prompts as context grows, and drops down to 40-50tok/s anyway making this config still moot for improvement on my Radeon card.
The only thing I can point to slowing me down is bandwidth of the card itself.
I am waiting to actually get my 5090 right now and I am betting that the 1700 Gbps of capacity will fix my prompt processing speeds. I don’t need full PCIe lane bandwidth to serve my house I just need to load the full model into vRAM and let the GPU do its thing.
Additional benefit to the external enclosure route is being able to migrate the inference between devices more easily. I can develop out the infrastructure then migrate the card to be hooked up to a shared node in the house with all the tools necessary for my family to take advantage of the privacy enhancement that comes with local inference.
How are you actually using the local model? I've played with Qwen 3.8 27b on ollama and the coding harnesses (Claude Code and OpenCode) seem to fail way more often then using the cloud models. And by fail, I mean the edits don't apply cleanly, it goes to add python code, but doesn't indent it properly, or the edit doesn't apply and so it tries again and again and eventually wipes out a different function then it "intended". It just gets really frustrating compared to the relative stability of Claude Cloud.
My use-case is only coding, every model sucks at writing good literature and there is no way around that (have had people try to debate me on this, but it's a taste thing, I have extensive English writing skills from my school years).
Prior to two weeks ago, I was just using Pi and Ollama.
I have tried my hand at putting together a few harnesses and I finally landed on what I like. Been working on this small app to handle running llama-server for me from any device that has the llama-cpp stack setup: https://github.com/SamInTheShell/loom
Qwen 3.8 is the first model I've been using that hasn't been having issues doing edit calls. Here are my llama server settings and GUFF that I use: https://gist.github.com/SamInTheShell/0bf838e8dc5093583b688e...
try using pi harness, hae not encountered these sort of problem myself also yuou can ask codex to look at the transcript and figure out the solutions to tool call failures that way
I should have mentioned in my prior comment, I didn't have any issues in Pi either.
try ninfer once you get your 5090 https://github.com/Neroued/ninfer
Thanks for sharing, I'm definitely trying it out after I get through my project milestones for the 5090. The README claims 700tok/s for Qwen 3.8 27b, that would be amazing, I'm only expecting an increase from my ~50tok/s on my Radeon to 200tok/s on the 5090.