Leanstral-1.5-119B-A6B-MLX-5bit

mvid/Leanstral-1.5-119B-A6B-MLX-5bit

Leanstral-1.5-119B-A6B-MLX-5bit is a image-text-to-text model from mvid. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for mlx. The card lists the license as apache-2.0.

5-bit MLX affine quantization of mistralai/Leanstral-1.5-119B-A6B, a 119B-parameter (A6B active) mixture-of-experts vision-language model tuned for Lean 4. Produced with mlxvlm.convert on Apple silicon.

Downloads · 30 days

8% of all-time downloads

All-time downloads

532

Public

Parameters

119B

83.4 GB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors83.4 GB · 100%

Parameter types

How the weights are stored.

U32118B · 99%

Try a prompt

Base models

mistralai/Leanstral-1.5-119B-A6B

Model card

From the Hugging Face model README

Leanstral-1.5-119B-A6B-MLX-5bit

Quantization recipe

Method: affine, group size 64, 5-bit.
Effective: ~5.58 bits/weight.
Router precision: the 36 MoE router gates (*.mlp.gate) are kept at 8-bit regardless of the target width, matching the reference recipe. This protects expert selection, which is disproportionately sensitive to quant noise.
Full precision: vision tower, multimodal projector, and lm_head are left unquantized.

Usage

pip install mlx-vlm
python -m mlx_vlm.generate \
  --model mvid/Leanstral-1.5-119B-A6B-MLX-5bit \
  --max-tokens 512 \
  --prompt "State and prove in Lean 4 that addition on the naturals is commutative."

MLA attention is expanded to full MHA in mlx-vlm's cache, so long-context runs are memory-heavy on this engine; for large contexts prefer a GGUF build under llama.cpp / LM Studio, which keeps the KV cache compressed.

Sibling quants

MLX-8bit · MLX-6bit · MLX-5bit · MLX-4bit · MLX-3bit

Task: Image-Text-to-Text
Library: mlx
Type: mistral3
License: apache-2.0
Created: Jul 7, 2026
Updated: Jul 9, 2026