Fix README: correct 228.7B params, add MLX Studio branding

README.md

@@ -23,16 +23,16 @@ pipeline_tag: text-generation
 
 # MiniMax-M2.7 JANG_3L
 
-**MiniMax M2.7 456B MoE — 3-bit mixed precision, 89 GB**
+**MiniMax M2.7 228B MoE — 3.08-bit mixed precision, 89 GB**
 
-Best balance of quality and size for Apple Silicon. Fits on 128 GB+ Macs.
+Best balance of quality and size — fits on 128 GB+ Macs.
 </div>
 
 > **Recommended: Run in [MLX Studio](https://mlxstudio.com)** for best experience including thinking mode support and optimized MoE inference.
 
 ## Important Settings
 
-MiniMax M2.7 requires specific inference settings:
+MiniMax M2.7 is an always-reasoning model. It thinks before answering on every prompt.
 
 | Setting | Value | Notes |
 |---------|-------|-------|
@@ -46,14 +46,16 @@ MiniMax M2.7 requires specific inference settings:
 | Metric | Value |
 |--------|-------|
 | Source | `MiniMaxAI/MiniMax-M2.7` (FP8 E4M3) |
-| Architecture | MoE (256 experts, top-8 active), GQA, partial RoPE |
+| Architecture | MoE (256 experts, top-8 active), GQA (48 heads / 8 KV), partial RoPE |
+| Total Parameters | 228.7B |
+| Active Parameters | ~1.4B per token |
 | Profile | JANG_3L (CRITICAL=8-bit, IMPORTANT=4-bit, COMPRESS=3-bit) |
 | Actual avg bits | 3.08 |
 | Model size | 89 GB |
-| Parameters | 456B total, ~46B active per token |
 | Format | JANG v2 (MLX-native safetensors, instant load) |
 | group_size | 128 (speed-optimized for 256 experts) |
-| Routing | Sigmoid + bias correction |
+| Routing | Sigmoid + bias correction (not softmax) |
+| QK-norm | Full vector RMSNorm |
 | Context | 192K tokens |
 
 ## JANG_3L Bit Allocation
@@ -62,28 +64,33 @@ MiniMax M2.7 requires specific inference settings:
 |------|-----------|------|
 | CRITICAL | Attention (Q/K/V/O), lm_head | 8 |
 | IMPORTANT | Embeddings | 4 |
-| COMPRESS | Expert MLP (w1/w2/w3) — 98.2% of params | 3 |
-| Passthrough | MoE router/gate (float16), norms | 16 |
+| COMPRESS | Expert MLP (w1/w2/w3) — 98%+ of params | 3 |
+| Passthrough | MoE router/gate (float16), norms, QK-norms | 16 |
 
-## MMLU Benchmarks
+JANG protects routing and attention at full precision while compressing the 256 expert MLPs — where MoE models are most tolerant of quantization. The router is kept at float16 (no quantization) for maximum routing precision.
 
-*Coming soon — benchmarks will be added after all profiles are converted and tested.*
+## MMLU Benchmarks (200q, 10 subjects, reasoning ON)
 
-## Why JANG
+*Coming soon — benchmarks in progress.*
 
-Standard MLX quantization on MiniMax M2.5 produced **completely broken output at ALL bit levels** (~25% MMLU = random guessing). JANG's mixed-precision approach is the **only working quantized MiniMax on Apple Silicon**, achieving 74% MMLU on M2.5. M2.7 results pending.
+## Why JANG for MiniMax
 
-## Other Quantizations
+Standard MLX quantization on MiniMax produces **completely broken output at ALL bit levels** (~25% MMLU = random guessing). JANG's mixed-precision approach is the **only working quantized MiniMax on Apple Silicon**.
 
-| Model | Profile | Size | Avg Bits | Status |
-|-------|---------|------|----------|--------|
-| JANG_2L | (8, 6, 2) | 63 GB | 2.10 | Ready |
-| **JANG_3L** (this) | **(8, 4, 3)** | **89 GB** | **3.08** | **Ready** |
-| JANG_4M | (8, 4, 4) | ~120 GB | ~4.1 | Converting |
+On M2.5, JANG_2L achieved **74% MMLU** vs MLX's 25% (random). M2.7 results pending.
+
+## All Quantizations
+
+| Model | Profile | Size | Avg Bits |
+|-------|---------|------|----------|
+| [JANG_2L](https://huggingface.co/JANGQ-AI/MiniMax-M2.7-JANG_2L) | (8, 6, 2) | 63 GB | 2.10 |
+| [JANG_3L](https://huggingface.co/JANGQ-AI/MiniMax-M2.7-JANG_3L) | (8, 4, 3) | 89 GB | 3.08 |
+| [JANG_4M](https://huggingface.co/JANGQ-AI/MiniMax-M2.7-JANG_4M) | (8, 4, 4) | 115 GB | 4.06 |
+| [JANG_6M](https://huggingface.co/JANGQ-AI/MiniMax-M2.7-JANG_6M) | (8, 6, 6) | 167 GB | 6.03 |
 
 ## Requirements
 
-- Apple Silicon Mac with 128+ GB unified memory
+- Apple Silicon Mac with 128 GB unified memory
 - MLX framework
 - [MLX Studio](https://mlxstudio.com) recommended
 
@@ -101,7 +108,7 @@ prompt = tokenizer.apply_chat_template(
     [{"role": "user", "content": "What is photosynthesis?"}],
     tokenize=False, add_generation_prompt=True
 )
-output = generate(model, tokenizer, prompt=prompt, max_tokens=500, sampler=sampler)
+output = generate(model, tokenizer, prompt=prompt, max_tokens=2048, sampler=sampler)
 print(output)
 ```
 
@@ -116,3 +123,4 @@ Quantized by Jinho Jang (eric@jangq.ai) using JANG Tools v2.4.1.
 ---
 
 *This model is provided for research and personal use. Users are responsible for ensuring their use complies with applicable laws and the MiniMax license.*
+