Skip to content

largha:150m-fp16

4 TagsUpdated 144M params512 context100+ languagesApache-2.0by Phocinae

Largha v1.1 (mmBERT-small, 144M): 0.398 on Bespoke Labs' public benchmark; 18 ms for five questions on an RTX 5090.

multilingual144m
ollaya run largha:150m-fp16 --preset triage "I was charged twice for my subscription this month and want a refund."

Details

  • graph17d1a327ea4f · 2 MBonnx · encoder · 144M · fp32
  • graphaf013f7b3912 · 3 MBonnx · encoder · 144M · fp16
  • weightsb6472511eea3 · 289 MBhuggingface.co/Phocinae/Phocinae-Largha-150M-v1/resolve/5300e85…/model.safetensors
  • tokenizer609d8f4c067c · 34 MBhuggingface.co/Phocinae/Phocinae-Largha-150M-v1/resolve/5300e85…/tokenizer/tokenizer.json
  • decision80dc4d8477bd · 441 B{"engine": "onnx", "family": "laya", "encoder": "encoder", "layout": "laya-markers-v1", …}
  • calibrationb9f584c723f3 · 129 B{"temperature": [0.866, 0.808, 0.662]}
  • licensef326fcf01f00 · 10 KBPhocinae-Largha-150M-v1 by Phocinae (https://huggingface.co/Phocinae/Phocinae-Largha-150M-v1)
  • archd2da87032dbb · 1 KBapplication/vnd.ollaya.arch
  • paramsa1d447ba672f · 21 Bprecision fp16

Every layer is checked against its sha256 when it is pulled. Weights and tokenizers download from the model author's Hugging Face repository at a pinned commit; Ollaya never re-hosts them.

Variant of

Largha readme and all models