Hugging Face's logo

batiai
/
MiniMax-M2.7-GGUF

Text Generation
GGUF
llama.cpp
minimax
quantized
apple-silicon
ollama
batiai
on-device
229b
imatrix
conversational
Model card Files
xet
 Community
hero775 commited on
Commit
a8a2475
·
verified ·
1 Parent(s): 4ddb16a

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +71 -114
README.md CHANGED
@@ -1,157 +1,114 @@
1
  ---
 
 
 
 
 
2
  license: other
3
  license_name: modified-mit
4
  license_link: https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE
5
- base_model: MiniMaxAI/MiniMax-M2.7
6
- base_model_relation: quantized
7
  tags:
8
- - gguf
9
- - quantized
10
- - llama.cpp
11
- - batiai
12
- - batiflow
13
- - moe
14
- - agentic
15
- - minimax
16
- library_name: gguf
17
  pipeline_tag: text-generation
 
18
  ---
19
 
20
- # MiniMax-M2.7 GGUF — Quantized by BatiAI
21
 
22
  <p align="center">
23
  <a href="https://flow.bati.ai"><img src="https://img.shields.io/badge/BatiFlow-macOS%20AI%20Automation-blue?style=for-the-badge&logo=apple" alt="BatiFlow"></a>
24
- <a href="https://ollama.com/batiai/minimax-m2.7"><img src="https://img.shields.io/badge/Ollama-batiai/minimax--m2.7-green?style=for-the-badge" alt="Ollama"></a>
25
  </p>
26
 
27
- > Optimized GGUF quantizations of **[MiniMaxAI/MiniMax-M2.7](https://huggingface.co/MiniMaxAI/MiniMax-M2.7)** for on-device AI on Mac.
28
- > Built and verified by [BatiAI](https://bati.ai) for [BatiFlow](https://flow.bati.ai) — free, unlimited, on-device AI automation for Mac. Just 5MB.
29
 
30
- ## About MiniMax-M2.7
31
 
32
- MiniMax-M2.7 is a **229B parameter MoE** agentic model (10B active per forward pass) released April 2026. It is the first model deeply participating in its own evolution:
33
-
34
- - **ELO 1495 on GDPval-AA** — highest among open-source models, surpassing GPT-5.3
35
- - **66.6% medal rate on MLE Bench Lite** (22 ML competitions), second only to Opus-4.6 and GPT-5.4
36
- - Built for **complex agent harnesses**, multi-round Word/Excel/PPT editing, and advanced productivity tasks
37
- - Self-evolution: internal version autonomously optimized its programming scaffold over 100+ rounds, achieving 30% performance improvement
38
 
39
  ## Quick Start
40
 
41
- ### Ollama (Recommended)
42
-
43
  ```bash
44
- # IQ3_XXS — 82GB, requires 128GB+ Mac
45
- ollama pull batiai/minimax-m2.7:q3
46
 
47
- # IQ4_XS — 114GB, requires 192GB+ Mac
48
- ollama pull batiai/minimax-m2.7:q4
49
 
50
- ollama run batiai/minimax-m2.7:q3
51
- ```
 
52
 
53
- ### llama.cpp
54
 
55
- ```bash
56
- # Download specific quant
57
- huggingface-cli download batiai/MiniMax-M2.7-GGUF MiniMaxAI-MiniMax-M2.7-IQ3_XXS.gguf --local-dir .
 
 
 
 
 
 
 
58
 
59
- # Run
60
- ./llama-cli -m MiniMaxAI-MiniMax-M2.7-IQ3_XXS.gguf -p "Hello"
61
- ```
62
 
63
- ### BatiFlow App
 
 
 
 
 
64
 
65
- [Download BatiFlow](https://github.com/batiai/batiflow-releases/releases/latest) → Settings → AI → Ollama → Select `batiai/minimax-m2.7`
66
 
67
- ## Available Quantizations
68
 
69
- | File | Quant Type | Size | Min RAM | Target Hardware |
70
- |------|-----------|------|---------|----------------|
71
- | `MiniMaxAI-MiniMax-M2.7-IQ3_XXS.gguf` | **IQ3_XXS** (imatrix) | **82GB** | 128GB | MacBook Pro M4 Max 128GB |
72
- | `MiniMaxAI-MiniMax-M2.7-IQ4_XS.gguf` | **IQ4_XS** (imatrix) | **114GB** | 192GB | Mac Studio M4 Ultra 192GB |
73
 
74
- All quants use **importance matrix** (imatrix) generated from wikitext-2-raw calibration data to preserve quality at low bit-widths. MoE models particularly benefit from imatrix quantization since expert utilization varies significantly.
 
 
 
 
 
 
75
 
76
- ## Why BatiAI Quantizations?
77
 
78
- | | BatiAI | Third-party (re-quantizations) |
79
  |---|---|---|
80
- | **Source** | Quantized directly from official MiniMax BF16 weights | Re-quantized from other GGUF files |
81
- | **imatrix** | Applied to all IQ quants (quality-optimized) | Often missing |
82
- | **Tested on** | Real Apple Silicon hardware | Untested on consumer hardware |
83
- | **Tool Calling** | Verified with BatiFlow's 57 tool functions | Often broken on MoE models |
84
- | **Korean** | Validated Korean text generation | Not tested |
85
- | **Metadata** | `general.author: BatiAI` signed | Varies |
86
 
87
  ## Technical Details
88
 
89
  - **Original Model**: [MiniMaxAI/MiniMax-M2.7](https://huggingface.co/MiniMaxAI/MiniMax-M2.7)
90
- - **Architecture**: MiniMax-M2 (MoE, 256 experts, 8 activated per token)
91
- - **Parameters**: 228.7B total, ~10B active per forward pass
92
- - **Context Window**: 196,608 tokens (192K native)
93
- - **License**: modified-MIT (commercial use permitted)
94
- - **Quantized with**: [llama.cpp](https://github.com/ggml-org/llama.cpp) (latest build, April 2026)
95
  - **Quantized by**: [BatiAI](https://bati.ai)
96
- - **GGUF Metadata**: `general.author: BatiAI`, `general.url: https://flow.bati.ai`
97
-
98
- ## How We Quantize
99
-
100
- ```
101
- MiniMaxAI/MiniMax-M2.7 official BF16 safetensors (~230GB)
102
- ↓ llama.cpp convert_hf_to_gguf.py
103
- BF16 GGUF (~457GB)
104
- ↓ llama-imatrix with wikitext-2 calibration (~6h CPU)
105
- imatrix.dat
106
- ↓ llama-quantize --imatrix
107
- IQ3_XXS (82GB) + IQ4_XS (114GB)
108
- ↓ BatiAI metadata signing
109
- Published GGUFs
110
- ```
111
-
112
- No third-party intermediaries. Direct from source, verified on real hardware.
113
 
114
  ## About BatiFlow
115
 
116
- <p align="center">
117
- <a href="https://flow.bati.ai"><strong>flow.bati.ai</strong></a>
118
- </p>
119
-
120
- [BatiFlow](https://flow.bati.ai) is a **macOS-native AI desktop automation app** — just 5MB, built with Swift.
121
-
122
- ### Why BatiFlow?
123
-
124
- | | |
125
- |---|---|
126
- | **Free & Unlimited** | On-device AI via Ollama — no API costs, no usage limits, no subscriptions |
127
- | **100% Private** | All data stays on your Mac. Nothing is sent to the cloud. |
128
- | **Ultra Lightweight** | Native macOS app, only 5MB. No Electron, no bloat. |
129
- | **Instant Setup** | Download → install Ollama → start automating. Under 5 minutes. |
130
-
131
- ### What It Does
132
-
133
- - **AI Command Bar** — type natural language, AI executes actions
134
- - **KakaoTalk, iMessage, Slack** — messaging automation
135
- - **Chrome browser** — navigate, click, fill, screenshot via CDP
136
- - **57 built-in tools** — calendar, notes, reminders, files, email, shell, and more
137
- - **Skill builder** — create reusable automations with YAML
138
- - **Multilingual** — Korean & English UI + AI responses
139
-
140
- <p align="center">
141
- <a href="https://github.com/batiai/batiflow-releases/releases/latest">
142
- <img src="https://img.shields.io/badge/Download-BatiFlow-blue?style=for-the-badge&logo=apple" alt="Download BatiFlow">
143
- </a>
144
- <a href="https://flow.bati.ai">
145
- <img src="https://img.shields.io/badge/Learn%20More-flow.bati.ai-green?style=for-the-badge" alt="Learn More">
146
- </a>
147
- </p>
148
 
149
  ## License
150
 
151
- This model is quantized from `MiniMaxAI/MiniMax-M2.7` and follows the original model's **modified-MIT license**. Commercial use is permitted per the original license terms.
152
-
153
- BatiAI quantization pipeline and metadata are provided under MIT License.
154
-
155
- ---
156
-
157
- **Benchmarks on Apple Silicon** (updated after verification): See [https://github.com/batiai/batiai-models](https://github.com/batiai/batiai-models) for reproducible benchmarks and quantization pipeline.
 
1
  ---
2
+ language:
3
+ - en
4
+ - ko
5
+ - ja
6
+ - zh
7
  license: other
8
  license_name: modified-mit
9
  license_link: https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE
 
 
10
  tags:
11
+ - gguf
12
+ - minimax
13
+ - quantized
14
+ - apple-silicon
15
+ - ollama
16
+ - batiai
17
+ - on-device
18
+ - 229b
19
+ base_model: MiniMaxAI/MiniMax-M2.7
20
  pipeline_tag: text-generation
21
+ library_name: llama.cpp
22
  ---
23
 
24
+ # MiniMax M2.7 GGUF — Quantized by BatiAI
25
 
26
  <p align="center">
27
  <a href="https://flow.bati.ai"><img src="https://img.shields.io/badge/BatiFlow-macOS%20AI%20Automation-blue?style=for-the-badge&logo=apple" alt="BatiFlow"></a>
28
+ <a href="https://ollama.com/batiai/minimax-m2.7"><img src="https://img.shields.io/badge/Ollama-batiai%2Fminimax--m2.7-green?style=for-the-badge" alt="Ollama"></a>
29
  </p>
30
 
31
+ > IQ3_XXS quantization of **MiniMaxAI/MiniMax-M2.7** (229B Dense) for on-device AI on Mac.
32
+ > Built and verified by [BatiAI](https://bati.ai) for [BatiFlow](https://flow.bati.ai).
33
 
34
+ ## Why MiniMax M2.7?
35
 
36
+ - **229B Dense** — one of the largest open models
37
+ - Outperforms GPT-5.3 on GDPval-AA (ELO 1495)
38
+ - Toolathon: 46.3% accuracy (global top tier)
39
+ - Agent Teams, complex Skills, dynamic tool search
40
+ - **Runs on a 128GB MacBook Pro** — no cloud needed
 
41
 
42
  ## Quick Start
43
 
 
 
44
  ```bash
45
+ ollama pull batiai/minimax-m2.7:iq3
46
+ ```
47
 
48
+ ## Available Quantizations
 
49
 
50
+ | Quant | Size | VRAM | M4 Max (128GB) | Recommended For |
51
+ |-------|------|------|----------------|----------------|
52
+ | **IQ3_XXS** | **82GB** | **104GB** | **36.7 t/s** | **128GB+ Mac** |
53
 
54
+ ## Benchmarks — MacBook Pro M4 Max (128GB)
55
 
56
+ | Metric | IQ3_XXS |
57
+ |--------|---------|
58
+ | **Token gen (short)** | **22.1 t/s** |
59
+ | **Token gen (long, 300 tokens)** | **36.7 t/s** |
60
+ | Prompt eval | 14.8 t/s |
61
+ | VRAM | 104 GB (97% GPU / 3% CPU) |
62
+ | Cold start | 42 seconds |
63
+ | Korean output | ✅ |
64
+ | Tool call JSON | ✅ |
65
+ | Basic math (2+2) | ✅ |
66
 
67
+ ### RAM Requirements
 
 
68
 
69
+ | Your Mac RAM | IQ3_XXS (82GB) |
70
+ |-------------|---------------|
71
+ | 64GB or less | ❌ Won't fit |
72
+ | 96GB | ⚠️ Heavy swap, unusable |
73
+ | **128GB** | **✅ 36.7 t/s** |
74
+ | 192GB+ | ✅ Fast, with headroom |
75
 
76
+ ## 229B on a Laptop
77
 
78
+ This is a 229B parameter dense model running entirely on-device — no cloud, no API, no costs. IQ3_XXS quantization compresses from 457GB (BF16) to 82GB while maintaining Korean, tool calling, and reasoning capabilities.
79
 
80
+ ## Model Comparison — Which BatiAI Model for Your Mac?
 
 
 
81
 
82
+ | Your Mac | Best Model | Speed |
83
+ |----------|-----------|-------|
84
+ | 16GB | `batiai/gemma4-e4b:q4` | 57 t/s |
85
+ | 24GB | `batiai/gemma4-26b:iq4` | 85 t/s |
86
+ | 36GB | `batiai/qwen3.5-35b:iq4` | 26.6 t/s |
87
+ | 48GB | `batiai/gemma4-31b:iq4` | 22.8 t/s |
88
+ | **128GB** | **`batiai/minimax-m2.7:iq3`** | **36.7 t/s** |
89
 
90
+ ## Why BatiAI Quantization?
91
 
92
+ | | BatiAI | Third-party (unsloth, etc.) |
93
  |---|---|---|
94
+ | **Source** | Quantized from [official MiniMax weights](https://huggingface.co/MiniMaxAI/MiniMax-M2.7) | Re-quantized from other GGUFs |
95
+ | **Tested on** | Real MacBook Pro M4 Max (128GB) | Often untested on consumer hardware |
96
+ | **Tool Calling** | ✅ Verified | Often untested |
97
+ | **Korean** | ✅ Validated | Not tested |
98
+ | **imatrix** | ✅ Calibrated for quality | Standard or none |
 
99
 
100
  ## Technical Details
101
 
102
  - **Original Model**: [MiniMaxAI/MiniMax-M2.7](https://huggingface.co/MiniMaxAI/MiniMax-M2.7)
103
+ - **Architecture**: Dense (229B params, all active)
104
+ - **License**: Modified-MIT
105
+ - **Quantized with**: [llama.cpp](https://github.com/ggml-org/llama.cpp)
 
 
106
  - **Quantized by**: [BatiAI](https://bati.ai)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
107
 
108
  ## About BatiFlow
109
 
110
+ [BatiFlow](https://flow.bati.ai) — free, on-device AI automation for Mac. 5MB app, 100% local, unlimited. 57+ built-in tools for calendar, notes, reminders, files, email, browser, messaging.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
111
 
112
  ## License
113
 
114
+ Quantized from [MiniMaxAI/MiniMax-M2.7](https://huggingface.co/MiniMaxAI/MiniMax-M2.7). License: **Modified-MIT**.