Spaces:

---

mlx-community

/

README

like 14

Running

Hi everyone,
I’ve been testing mlx-community/IQuest-Coder-V1-14B-Thinking and mlx-community/IQuest-Coder-V1-14B-Instruct on Apple Silicon (M4 MacBook Air 24GB).
From my experiments, 6-bit quantization seems like a very strong sweet spot for this class of 14B coding models on 24GB unified memory systems.
8-bit works well but uses noticeably more memory, especially with 4K–8K context.
4-bit is more memory efficient but I’ve seen some degradation in reasoning quality for coding tasks.
6-bit feels like a balanced middle ground (quality vs memory vs speed), especially for MLX workloads.
Because of this, I was wondering:
👉 Would it be possible to generate and publish official MLX 6-bit variants for both:
IQuest-Coder-V1-14B-Thought
IQuest-Coder-V1-14B-Instruct
It would be very useful for MacBook users (especially 24GB M-series devices) to have an optimized 6-bit option available in the community repo.
If someone has already tested this or has a recommended quantization config (group size / mode), I’d really appreciate your input.
Thanks!