Mistral-Small-4-119B-2603-OptiQ-2bit

Built with mlx-optiq, the MLX-native toolkit to quantize, fine-tune, and serve LLMs locally on Apple Silicon (no PyTorch, no cloud). Try the Lab · All OptiQ quants · Docs

Supported loaders: mlx-optiq (text, vision, and MTP) and stock mlx-lm (text). Other front-ends load MLX weights through their own stack, so support there depends on that stack rather than on these files.

An OptiQ mixed-precision MLX quant of Mistral-Small-4-119B-2603, a 128-routed-expert sparse-MoE flagship. This is an extreme 2-bit build meant to run a 119B model on a 36 GB Mac.

• Mixed 2/4-bit static build — per-layer bit-widths assigned to a 2.5 target bits-per-weight: 352 projections at 4-bit, 82 at 2-bit.
• 239 GB bf16 to 43 GB on disk (5.6x).
• Runs on a 36 GB Mac. OptiQ streams the routed experts off SSD on demand, so the resident footprint is ~9-12 GB while the full model never materializes.

Requirements

Needs optiq >= 0.4.4, which ships the vendored mistral4 decoder (the 128-expert MoE with MLA attention). Stock mlx-lm and older OptiQ load the base repo's mistral3 wrapper, which falls back to llama and drops every expert. Do not load this repo that way.

pip install -U optiq

Running it

This is a large sparse MoE, so serve it with SSD expert streaming (auto-detected):

optiq serve --model mlx-community/Mistral-Small-4-119B-2603-OptiQ-2bit

optiq serve installs the mistral4 decoder, detects the streamable MoE, and loads it via optiq.runtime.moe_stream so the routed experts are read from disk per token instead of held in RAM. Use the OpenAI-compatible endpoint at http://localhost:8000/v1.

No Capability Score by design

Extreme 2-bit quants of 100B+ models ship with a coherence demo instead of a benchmark suite. At 2-bit, this quant writes a complete, self-contained Flappy Bird that actually plays: