JANGQ-AI/MiniMax-M2.5-JANG_3L

MLX Studio — native JANG support with reasoning

---

MiniMax M2.5 (227B-A21B) — JANG_3L (3.08-bit) — Reasoning

JANG — Jang Adaptive N-bit Grading | The GGUF Equivalent for MLX

     

JANG is fully open-source. Quantization engine and full commit history: github.com/jjang-ai/jangq. Created by Jinho Jang.

---

Supported apps: MLX Studio (full native support) and oMLX (PR #364). LM Studio, Ollama, and Inferencer do not yet support JANG.

---

Why JANG models?

Tools like mlx-lm, oMLX (oQ), and others can quantize models — but shipping a tool is the easy part. JANG models come from hundreds of hours of per-architecture testing: finding which layers break at which bit depths, which MoE routing survives quantization, which models need bfloat16 to avoid NaN. We don't just quantize — we convert, verify, benchmark, and publish every model with tested scores. No other project in the MLX ecosystem publishes pre-tested quantized models at this scale.

---

Results: 93.5% MMLU (200 Questions, Smart Two-Pass)

Subject	Score
Abstract Algebra	11/20 (55%)
Anatomy	19/20 (95%)
Astronomy	20/20 (100%)
College CS	20/20 (100%)
College Physics	19/20 (95%)
HS Biology	20/20 (100%)
HS Chemistry	19/20 (95%)
HS Mathematics	20/20 (100%)
Logical Fallacies	19/20 (95%)
World Religions	20/20 (100%)
Total	187/200 (93.5%)

Pass 1 (no-thinking): 158/200 (79.0%) | Pass 2 (reasoning retry): +29 recovered

JANG vs MLX — MiniMax M2.5

Model	MMLU	Size	Speed	Notes
JANG_3L (this model)	93.5%	82 GB	41 tok/s	5 subjects at 100%
JANG_2L	74.0%	63 GB	48 tok/s	Smallest working MiniMax
MLX 4-bit	26.5%	91 GB	~50 tok/s	Broken — random answers
MLX 3-bit	24.5%	69 GB	—	Broken — random answers
MLX 2-bit	25.0%	46 GB	—	Broken — random answers

MLX is broken on MiniMax at ALL bit levels (~25% = random chance). JANG is the ONLY working quantization for MiniMax M2.5 on Apple Silicon.

Key Features

• 93.5% MMLU — five subjects at 100%
• 41 tok/s generation on M3 Ultra
• 82 GB on disk — fits 96+ GB Macs
• 227B total / 21B active — 256 MoE experts, top-8 routing
• Reasoning mode: <think>...</think> step-by-step reasoning
• Sigmoid + bias routing: MiniMax-specific MoE (not softmax)
• FP8 source with block-wise 128x128 scales

Important Notes

• Temperature must be 1.0 — greedy decoding (temp=0) causes infinite thinking loops on MiniMax
• Tokenizer: Known-good tokenizer included (mlx_lm.convert corrupts MiniMax tokenizer)

Architecture

227B total parameters, 21B active per token
- 64 layers, all MoE (256 experts, top-8 routing)
- Sigmoid + bias expert routing (non-normalized)
- GQA attention: 48 heads, 8 KV heads
- FP8 E4M3 source with block-wise scales

Install

pip install jang[mlx]

---

Created by Jinho Jang — jangq.ai — @dealignai