← sources

Best local models for coding and chat on a 96 GB M3 Ultra Mac Studio?

source: reddit (reddit:LocalLLaMA)published 2026-10-19 16:14ingested 2026-10-20 02:39

retained text

I’ve been running Qwen3.8-27B and have been pretty impressed, especially for coding. I’m wondering what else people are finding worth running on this hardware.

My current setup:

• M3 Ultra Mac Studio with 96 GB unified memory.

• llama.cpp with Metal, served through llama-swap (only one model loaded at a time)

• Coding: Qwen3.8-27B, Unsloth Q8\_0, with MTP speculative decoding and a configured 262,144-token context.

• Chat: Qwen3-Coder-Next, Unsloth UD-Q6\_K\_XL, with a configured 131,072-token context.

• Q8\_0 KV cache for both.

I haven’t been as impressed with Coder-Next for coding in my setup, so I’ve mostly been using it for chat instead.

Are there newer models worth trying over these, or different quants that would make better use of the machine? I’m interested in coding with tool use as well as general chat; they don’t need to be the same model.

If you’re running similar hardware, I’d love to hear the exact model, quant, backend, and context size you’re using, plus what performance you’re seeing.

  
\[EDIT\]: This machine is literally only used as an LLM inference server. So I can really fill all the memory up with a model. I've raised the OS memory limit to 86Gb, thinking about pushing it further...

document fetched 2026-10-20 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1xa1715/best_local_models_for_coding_and_chat_on_a_96_gb