Hugging Face's logo Hugging Face
  • Models
  • Datasets
  • Spaces
  • Buckets new
  • Docs
  • Enterprise
  • Pricing

  • Log In
  • Sign Up

Tinman-Lab
/
Tinman-SmolOmni-MLA-500M

Transformers
English
mla-attention
multi-head-latent-attention
flow-matching
rectified-flow
on-device
efficient-attention
smol-scale
research
proof-of-concept
Model card Files Files and versions
xet
Community

Instructions to use Tinman-Lab/Tinman-SmolOmni-MLA-500M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

  • Libraries
  • Transformers

    How to use Tinman-Lab/Tinman-SmolOmni-MLA-500M with Transformers:

    # Load model directly
    from transformers import AutoModel
    model = AutoModel.from_pretrained("Tinman-Lab/Tinman-SmolOmni-MLA-500M", dtype="auto")
  • Notebooks
  • Google Colab
  • Kaggle
Tinman-SmolOmni-MLA-500M
1.17 GB
Ctrl+K
Ctrl+K
  • 1 contributor
History: 32 commits
TinmanLabSL's picture
TinmanLabSL
HONEST README: Research POC with full prod-readiness roadmap, costs, verified architecture
cc53efa verified 15 days ago
  • benchmarks
    Upload benchmarks/eval_500m_asr_results.json 15 days ago
  • stage2_final
    Stage 2 FINAL: 2000 steps, loss 6.45→1.43 (78% reduction), flow head trained 17 days ago
  • .gitattributes
    1.52 kB
    initial commit 17 days ago
  • README.md
    6.41 kB
    HONEST README: Research POC with full prod-readiness roadmap, costs, verified architecture 15 days ago
  • benchmark_500M.json
    640 Bytes
    Benchmarks: 46.9% KV reduction, 15.9K tok/s, 1239MB VRAM 17 days ago
  • config.json
    1.33 kB
    Upload config.json 15 days ago
  • needle_haystack_500M.json
    2.17 kB
    Needle-in-haystack: 512-2048 tokens, no OOM, all depths pass 17 days ago