JANG is fully open-source. Quantization engine and full commit history: github.com/jjang-ai/jangq. Created by Jinho Jang.
First Mistral Small 4 (119B) on Apple Silicon. MLA attention + 128 MoE experts + Pixtral VLM. 5x faster prefill than MLX Community 4-bit.
Reasoning mode: Set reasoning_effort to "high" for step-by-step reasoning with [THINK]...[/THINK] tags.
Speed Comparison — JANG vs MLX Community
| Model | Size | Gen tok/s | Prefill tok/s | RAM | Fits On |
|---|---|---|---|---|---|
| JANG_2L (this model) | 30 GB | 82 | 216 | 40 GB | 48 GB Macs |
| JANG_4M | 57 GB | 80 | 202 | 68 GB | 96+ GB Macs |
| JANG_6M | 84 GB | 74 | 160 | 95 GB | 128+ GB Macs |
| MLX Community 4-bit | 63 GB | 84 | 43 | 68 GB | 96+ GB Macs |
- 5x faster prefill (216 vs 43 tok/s)
- Half the size (30 GB vs 63 GB) at comparable generation speed
- Benchmarked on M3 Ultra 256 GB with bfloat16 compute
Key Features
- 82 tok/s generation on M3 Ultra — matches MLX 4-bit at half the size
- 30 GB on disk, 40 GB peak RAM — fits 48 GB Macs (M4 Pro, M2/M3 Max)
- Vision (VLM): Pixtral encoder, 1540px max, processes images
- Reasoning mode: [THINK]...[/THINK] step-by-step reasoning
- Code generation: Complete functions with docstrings and optimized logic
- Math: Step-by-step calculations with distributive property
- 119B total / 6B active per token — MLA attention + 128 MoE experts
Architecture
JANG_2L Bit Allocation
| Tensor Type | Bits | Purpose |
|---|---|---|
| Attention (q/k/v/o projections) | 8 | Critical — preserves MLA precision |
| Embeddings, lm_head | 8 | Critical — token representation |
| MoE gate (router) | 16 | Float16 passthrough — routing precision |
| Shared experts | 6 | Important — always active |
| Routed experts (128) | 2 | Compressed — many experts = redundancy |
| Norms, biases | full | Float — tiny tensors, keep exact |
Results: 94.0% MMLU at 2-bit (200 Questions, Reasoning Mode)
| Subject | Score |
|---|---|
| Abstract Algebra | 13/20 (65%) |
| Anatomy | 20/20 (100%) |
| Astronomy | 20/20 (100%) |
| College CS | 20/20 (100%) |
| College Physics | 19/20 (95%) |
| HS Biology | 19/20 (95%) |
| HS Chemistry | 20/20 (100%) |
| HS Mathematics | 18/20 (90%) |
| Logical Fallacies | 19/20 (95%) |
| World Religions | 20/20 (100%) |
| Total | 188/200 (94.0%) |
Five perfect 100% subjects. 119B intelligence in 30 GB at 82 tok/s.
Requirements
- MLX Studio for native JANG support with reasoning
- Or: - Apple Silicon Mac with 48+ GB unified memory
Install
\
Created by Jinho Jang — jangq.ai — @dealignai