Hugging Face's logo

Spaces:
baa-ai
/
README
Running

App Files Community
tomkay commited on
Commit
ffc4e69
·
verified ·
1 Parent(s): b2f0907

Update org card: focus on models, remove paper references

Browse files
Files changed (1) hide show
  1. README.md +38 -30
README.md CHANGED
@@ -9,45 +9,53 @@ pinned: false
9
 
10
  <div style="text-align: center; padding: 2em 1em;">
11
 
12
- <center><a href="https://baa.ai"><img src="https://raw.githubusercontent.com/baa-ai/MINT/main/baa-logo-dark.svg" alt="baa.ai logo" width="128" height="128" style="margin-bottom: 1em;"></a>
13
- </center>
14
- <h1><a href="https://baa.ai" style="text-decoration: none; color: inherit;">baa.ai</a></h1>
15
  <h2>Smaller. Smarter. Sovereign.</h2>
16
  <h3>Making frontier models run anywhere</h3>
17
 
18
  <p style="max-width: 640px; margin: 1em auto; line-height: 1.6;">
19
- We build open tools for efficient AI deployment. Our research focuses on quantization methods that preserve model quality while dramatically reducing hardware requirements — bringing 400B+ parameter models to a single machine.
20
  </p>
21
 
22
- <p style="margin-top: 2em;">
23
- <a href="https://baa.ai">Website</a>&ensp;·&ensp;<a href="https://baa.ai/articles">Research</a>&ensp;·&ensp;<a href="https://github.com/baa-ai/MINT">GitHub</a>&ensp;·&ensp;<a href="https://huggingface.co/spaces/baa-ai/MINT">MINT Space</a>
24
  </p>
25
 
26
  </div>
27
 
28
  ---
29
 
30
- ### Research
31
-
32
- Read the full paper: **[MINT: Compute-Optimal Data-Free Mixed-Precision Quantization for LLMs](https://baa.ai/articles/24-mint-paper.html)**
33
-
34
- * **Papers**
35
- * [MINT: Compute-Optimal Data-Free Mixed-Precision Quantization for LLMs](https://baa.ai/articles/24-mint-paper.html)
36
- * [SWAN: Data-Free Mixed-Precision Quantization via Multi-Metric Sensitivity Analysis](https://baa.ai/articles/07-swan-data-free-mixed-precision.html)
37
- * **Benchmark Results**
38
- * [MINT Benchmark Results: 7 Models, 40,000+ Questions, One Winner](https://baa.ai/articles/38-mint-benchmark-results.html)
39
- * [Beyond Perplexity: Downstream Benchmarks Confirm MINT Beats All Quantization Strategies](https://baa.ai/articles/35-downstream-benchmarks-confirm.html)
40
- * [Eight Things Our Benchmarks Reveal That Nobody Expected](https://baa.ai/articles/39-eight-things-benchmarks-reveal.html)
41
- * **Key Findings**
42
- * [When Data-Free Beats the Gold Standard: MINT vs GPTQ](https://baa.ai/articles/27-mint-beats-gptq-data-free.html)
43
- * [The Compression Variable Everyone Ignored](https://baa.ai/articles/25-group-size-primary-quality-lever.html)
44
- * [Mean Perplexity Is Lying to You](https://baa.ai/articles/33-mean-perplexity-is-lying.html)
45
- * [When Better Is Worse: What We Learned by Trying to Improve Our Quantizer](https://baa.ai/articles/40-when-better-is-worse-quantizer.html)
46
- * **Technical Deep Dives**
47
- * [MLX Quantization on Apple Silicon: How MINT Turns a Mac into a Model Compression Lab](https://baa.ai/articles/32-mlx-quantization-apple-silicon.html)
48
- * [How to Fit Massive Models onto Tiny Memory Footprints](https://baa.ai/articles/28-budget-targeted-quantization-mckp.html)
49
- * [Eight Measurements per Tensor: The End of Single-Point Sensitivity](https://baa.ai/articles/30-rate-distortion-curves.html)
50
- * [How to Quantize a Model with 256 Experts](https://baa.ai/articles/31-expert-grouped-allocation-moe.html)
51
- * [One Number That Prevents Catastrophic Quantization](https://baa.ai/articles/26-sqnr-9db-safety-threshold.html)
52
-
53
- Browse all research at [baa.ai/articles](https://baa.ai/articles)
 
 
 
 
 
 
 
 
 
 
 
9
 
10
  <div style="text-align: center; padding: 2em 1em;">
11
 
12
+ <h1>🐑 baa.ai</h1>
 
 
13
  <h2>Smaller. Smarter. Sovereign.</h2>
14
  <h3>Making frontier models run anywhere</h3>
15
 
16
  <p style="max-width: 640px; margin: 1em auto; line-height: 1.6;">
17
+ We publish high-quality mixed-precision quantized models for Apple Silicon and GGUF. Our models use MINT allocation — a cutting-edge per-tensor sensitivity analysis that assigns optimal bit-widths under your exact memory budget.
18
  </p>
19
 
20
+ <p style="margin-top: 1.5em;">
21
+ <a href="https://baa.ai">🌐 Website</a>&ensp;·&ensp;<a href="https://github.com/baa-ai/MINT-UI">🛠 MINT-UI (create your own)</a>
22
  </p>
23
 
24
  </div>
25
 
26
  ---
27
 
28
+ ### 🤗 Models
29
+
30
+ All models are published as **MLX** (Apple Silicon) and **GGUF** (cross-platform) formats.
31
+
32
+ | Model Family | Sizes Available | Format |
33
+ |-------------|----------------|--------|
34
+ | **Qwen3.5-397B** | 220GB, 224GB | MLX, GGUF |
35
+ | **Qwen3.5-122B** | 52GB, 128GB, 154GB | MLX, GGUF |
36
+ | **Qwen3.5-35B** | 15–51GB (8 variants) | MLX, GGUF |
37
+ | **Llama-4-Maverick (402B)** | 407GB | MLX, GGUF |
38
+ | **Llama-4-Scout (109B)** | 117GB | MLX, GGUF |
39
+ | **Llama-3.1/3.3-70B** | 47GB | MLX |
40
+ | **MiniMax-M2.5 (229B)** | 179GB | MLX, GGUF |
41
+ | **Nemotron-120B** | — | MLX, GGUF |
42
+ | **GLM-4.7-Flash** | 16GB | MLX |
43
+ | **Qwen3-30B/8B** | 16GB, 6GB | MLX |
44
+
45
+ ### 🚀 Create Your Own
46
+
47
+ Use [**MINT-UI**](https://github.com/baa-ai/MINT-UI) to quantize any model to your exact memory target:
48
+
49
+ ```bash
50
+ pip install mint-ui
51
+ mint-ui
52
+ ```
53
+
54
+ - ⚡ Analysis completes in under 60 seconds
55
+ - 🎯 Specify any memory budget (e.g., "fit in 24 GB")
56
+ - 🧠 No calibration data required
57
+ - 💻 Runs on any Mac — even 32 GB machines can analyze 400B+ models
58
+
59
+ <p style="margin-top: 2em; text-align: center;">
60
+ <a href="https://baa.ai">baa.ai</a>
61
+ </p>