Skip to content

d3:edge

6 TagsUpdated 0.6B params8192 contextEnglishApache-2.0by vLLM Semantic Router

Qwen3.5-0.8B cut to 12 layers, the fastest: 28 ms for five questions on an RTX 5090.

0.6b
ollaya run d3:edge --preset triage "I was charged twice for my subscription this month and want a refund."

Details

  • graphd1e4addcc1c7 · 4 MBonnx · d3 · 0.6B · fp32
  • weights91bb8618450a · 1.2 GBhuggingface.co/vllm-sr/d3-edge/resolve/c454ee3…/model-00001-of-00001.safetensors
  • weightsde33d5dc795b · 1 MBhuggingface.co/vllm-sr/d3-edge/resolve/c454ee3…/readout.safetensors
  • tokenizer06b9509352d2 · 20 MBhuggingface.co/bespokelabs/Bespoke-Nimble-9B-v2/resolve/4b8c04d…/tokenizer.json
  • decisionb5058688f3fd · 5 KB{"engine": "onnx", "family": "d3", "encoder": "", "layout": "d3-code-readout-v1", …}
  • calibration0f89e51fde90 · 133 B{"temperature": [1.0, 1.0, 1.0]}
  • license3d7ec11263ac · 10 KBDecision 3.0 (d3) by the vLLM Semantic Router contributors (https://huggingface.co/vllm-sr)

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.

Decision 3.0 (d3) readme and all models