Qwen3.6-27B DSpark draft head (clean retrain)
DSpark speculative-decoding draft head for Avesed/Qwen3.6-27B-W4A16,
trained fully on-policy on the target's own responses.
Accepted length (num_speculative_tokens=7)
| task | accept_len |
|---|---|
| gsm8k | 6.55 |
| humaneval | 3.84 |
| math | 4.11 |
| chat | 2.54 |
Multilingual (incl. Chinese) generation verified clean — this revision was retrained on a freshly re-verified (sha256-checked) target checkpoint, fixing a non-English degeneration present in the earlier revision.
Serve (vLLM)
--speculative-config '{"method":"dspark","model":"Avesed/Qwen3.6-27B-DSpark","num_speculative_tokens":7}'
- Downloads last month
- 53
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support