ะ—ะฐะฟัƒัั‚ะธั‚ัŒ ั‡ะตั€ะตะท Basalt Run with Basalt bas-alt.vercel.app Basalt v0.11.0

โšก ะ—ะฐะฟัƒัั‚ะธั‚ัŒ ะฒ Basalt v0.11.0 ยท bas-alt.vercel.app (macOS 27+ ยท Swift 6.4 ยท Apple Silicon, ะฑะตะท Python):

basalt serve --model goodmai/Qwen3.8-27B-3bit-PVTuned-MLX --rest

Qwen3.8-27B-3bit-PVTuned-MLX

3-bit MLX affine checkpoint of Qwen3.8-27B refined with Basalt PV-Tuning against top-64 teacher log-probabilities with monotone held-out KL rollback.

Measured Quality & Footprint (Apple M4 Pro 24 GB)

Metric Measured Value
Method 3-bit MLX Affine + Basalt PV-Tuning (#410 / #486)
Weights Footprint 11.77 GB
WikiText-2 16k PPL (4ร—4096) 4.652 (vs 3-bit RTN 4.702; 4-bit GPTQ teacher 4.130)
System One Accuracy (N=300) 293 / 300 (97.7%)
Reproduce Command basalt models convert mlx-community/Qwen3.8-27B-GPTQ-4bit --quantize affine:3:64 --experimental-pv-tune --out ./out
Downloads last month
30
Safetensors
Model size
27B params
Tensor type
U32
ยท
BF16
ยท
MLX
Hardware compatibility
Log In to add your hardware

3-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for goodmai/Qwen3.8-27B-3bit-PVTuned-MLX

Base model

Qwen/Qwen3.5-27B
Quantized
(248)
this model

Collection including goodmai/Qwen3.8-27B-3bit-PVTuned-MLX