🦖 Osaurus Exclusive — a compact JANG-quantized MiniMax-M3 coder (coding · agentic · multimodal) for Apple Silicon.
> ⚠️ **Requires vMLX engine v1.5.67+.** > This is a **JANG-format** model (JANG affine + **AWQ** quant, **REAP** expert pruning, MiniMax-M3 MSA/Lightning-Indexer runtime). It will **NOT** load with `transformers`, `vLLM`, or generic MLX loaders — it runs on the vMLX engine (ships in **Osaurus**). ## What is a JANG model? **JANG** is vMLX's quantization + packing format: mixed-precision affine quant with per-projection bit widths + **AWQ** activation-aware scaling + **REAP** expert pruning, via a `jang_config.json`. Weights stay quantized in GPU memory and load through vMLX's JANG loader. The format + the M3 runtime are vMLX-specific, so it **runs only on vMLX 1.5.67 or newer.** ## Highlights - **Smallest M3 coder — ~84 GB** (the compact Osaurus build). - **REAP45:** keep **70/128** routed experts (45% pruned). - **All-2-bit routed experts + AWQ** (gate/up 2-bit AWQ-scaled, down 2-bit); attention 8-bit, shared experts 6-bit, embeddings 6-bit, lm_head 8-bit, Lightning Indexer FP16. - **Multimodal (vision) kept.** - Calibration: Vera (agentic-coder) + GSM8K; "floor" recipe keeps the most-salient coding experts. ## Run it - In **Osaurus** / vMLX 1.5.67+: pick this model, Start, then chat. - CLI: `vmlx-engine serve OsaurusAI/MiniMax-M3-Coder-Small --reasoning-parser minimax_m3 --tool-call-parser minimax_m3` ## Attribution - Base model: **MiniMaxAI/MiniMax-M3** · Pruning: **REAP** (Cerebras, arXiv:2510.13999) - **Vera calibration + testing: [@hornsman1](https://huggingface.co/hornsman1) (hornsan1 on GitHub)** · math calibration: GSM8K - Quantization & runtime: **JANG / vMLX** · Distributed via **Osaurus**