How is this easier than using lmstudio or omlx or whatever your favourite runtime is ?
It’s maybe a bit interesting that Jetbrains are making moves to integrate with local models more closely, but I think claiming this is “easier” is incorrect for most users.
The company also said that it chose Qwen3.6 over the newer Qwen3.8 (released earlier this month) because the latter runs slower on "today's Macs" since it needs reasoning enabled.
Why are they tying the client with the AI model instead of using OpenAI or another popular http format? Running llama.cpp locally is about the right level of complexity for most doing local AI.
It seems to me that Qwen3.6-35B-A3B is still the local LLM leader, because 27B in either of the latest releases is just too slow to be usable compared to OpenCode Zen free models, or OpenRouter free models.
I get up to 90 tokens / second with Jundot/Qwen3.6-35B-A3B-oQ6-mtp, on a M3 Max with 64GB. MoE so it is not a dense model, but that means it runs faster (also, mtp helps). It is a 30GB model, but you should be able to load it, otherwise try the 4-bit quant instead of the 6-bit quant, don't bother quanting your KV Cache (don't enable turboquant in oMLX), since that will slow you down.
I'm not sure what that means on a M5 max, definitely faster, I don't know if it really plays into the strengths of the new chip design though.
As far as I know, if you want MTP you need to serve it with either MTPLX or oMLX. oMLX is more stable and user friendly.
I’m using Goose but have tested Codex and OpenClaw among others. Codex worked fine but I couldn’t get web search to work with Qwen. OpenClaw I could get web search to work but its system prompt eats a lot of context tokens. Goose works for web search if you use SearchXNG, and has fairly slim system prompts. I’m still evaluating DeepSeek Harness and might move to that once it’s more stable. I tried other more obscure options but they had some flaws that made me settle on goose.
How is this easier than using lmstudio or omlx or whatever your favourite runtime is ?
It’s maybe a bit interesting that Jetbrains are making moves to integrate with local models more closely, but I think claiming this is “easier” is incorrect for most users.
Curious why they're using 3.6-27B and not 3.8-27B which is competitive with Opus 4.6 (https://huggingface.co/Qwen/Qwen3.8-27B)
From the article:
The company also said that it chose Qwen3.6 over the newer Qwen3.8 (released earlier this month) because the latter runs slower on "today's Macs" since it needs reasoning enabled.
But that's incorrect, reasoning can be disabled via the template.
I haven’t found that to be true with LM Studio and Qwen3.8. Works fine without reasoning.
They’re also targeting 64GB M5 Pro and up as “today’s Macs” which perform fine with reasoning enabled.
Why are they tying the client with the AI model instead of using OpenAI or another popular http format? Running llama.cpp locally is about the right level of complexity for most doing local AI.
It seems to me that Qwen3.6-35B-A3B is still the local LLM leader, because 27B in either of the latest releases is just too slow to be usable compared to OpenCode Zen free models, or OpenRouter free models.
Sad a Qwen3.8-35B-A3B model wasn't released.
Also, I think this is just an ad.
Anything good to run on Mac M5 Max with 48GB? is this even worth trying? so far I found the responses so slow compared to the paid subscriptions...
I get up to 90 tokens / second with Jundot/Qwen3.6-35B-A3B-oQ6-mtp, on a M3 Max with 64GB. MoE so it is not a dense model, but that means it runs faster (also, mtp helps). It is a 30GB model, but you should be able to load it, otherwise try the 4-bit quant instead of the 6-bit quant, don't bother quanting your KV Cache (don't enable turboquant in oMLX), since that will slow you down.
I'm not sure what that means on a M5 max, definitely faster, I don't know if it really plays into the strengths of the new chip design though.
Thanks! I have to try that! Might be tight! Can it run in Claude Code? Are you loading it with Ollama?
As far as I know, if you want MTP you need to serve it with either MTPLX or oMLX. oMLX is more stable and user friendly.
I’m using Goose but have tested Codex and OpenClaw among others. Codex worked fine but I couldn’t get web search to work with Qwen. OpenClaw I could get web search to work but its system prompt eats a lot of context tokens. Goose works for web search if you use SearchXNG, and has fairly slim system prompts. I’m still evaluating DeepSeek Harness and might move to that once it’s more stable. I tried other more obscure options but they had some flaws that made me settle on goose.
Nice, I will give it a try! Thanks!
Uh...
claude> dl & config best vers of Qwen 3.8 for my system