⚠️ JANG-format model — runs on Osaurus. This uses the JANG quantization format (mixed-precision affine + AWQ + REAP expert pruning) and loads through Osaurus's native Swift runtime. It will NOT load with
transformers,vLLM, or generic MLX loaders.
What is a JANG model?
JANG is a mixed-precision quantization + packing format — per-projection affine bit widths + AWQ activation-aware scaling + REAP expert pruning — described by a jang_config.json. Weights stay quantized in GPU memory. Osaurus loads it through its native Swift JANG runtime on Apple Silicon.
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.
Benchmarks
- HumanEval: pass@1 = 100% (82/82, scrambled-half adaptive eval, seed 42; 0 failures, 0 escalations).
- Despite 45% expert pruning + all-2-bit routed experts, coding accuracy holds at 100% — the REAP45 keep-set is a subset of the larger M3-Coder builds' proven coding experts, so coding capability is preserved while the model shrinks to ~84 GB.
Run it
Load it in Osaurus (Apple Silicon) — it runs on Osaurus's native Swift JANG runtime.
Attribution
- Base model: MiniMaxAI/MiniMax-M3 · Pruning: REAP (Cerebras, arXiv:2510.13999)
- Vera calibration + testing: @hornsman1 (hornsan1 on GitHub) · math calibration: GSM8K
- Quantization: JANG · Runtime & distribution: Osaurus