🧠 Qwen3-1.7B-Thinking-Distil

Extended reasoning distilled from Qwen3-30B-A3B-Thinking into a 1.7B student (~2.03B params, BF16, 40K context) via SFT on longwriter-6k. The model deliberates before answering — its chain of thought streams into the collapsible Reasoning panel, with the final answer below it.

Model card · Apache-2.0 · running on ZeroGPU

System prompt

Max new tokens

The model was trained at 4096 max seq — give long reasoning room.

256 4096

Temperature

0.6–0.8 activates the deliberation patterns without losing coherence.

0 1.5

Top-p

0.05 1

Top-k

1 100

Repetition penalty

1.1–1.2 keeps long generations out of reasoning loops.

1 1.5

Enable thinking

Chatbot