Small dense models can inherit the multi-step reasoning style of frontier-scale teachers — and that is the core experiment behind this release. Qwen3.8-9B Distill compresses Qwen3.8 2.4T's chain-of-thought behaviour into a full-parameter 9B student by fine-tuning on roughly 70,000 curated teacher traces, producing a deployable model that retains advanced reasoning on many benchmarks while running on a single GPU.
Key Capabilities
- Distilled chain-of-thought: the student reproduces teacher-style CoT outputs (answers begin with a learned
<think>block), which helps on multi-step reasoning and knowledge-intensive prompts. - Benchmark profile: large gains on MMLU (flexible-extract acc ~0.751 vs base 0.546, +0.205) while showing a small drop on GSM8K exact-match metrics; this indicates strong cross-task transfer for subject-level knowledge but occasional formatting/precision differences on some arithmetic tasks.
- Deployment-ready features: native 262,144-token context, native function-calling compatible with Qwen3.5 conventions, and full-parameter fine-tune (not an adapter), plus common quantizations (GGUF) and runtimes support (Transformers, vLLM, llama.cpp tooling).
Who it's for & Trade-offs
Great fit if you need a single‑GPU model that preserves large‑model reasoning style for tasks like multi-subject QA, code reasoning, or benchmark evaluation, and you value long context and native function-calling. Look elsewhere if strict per-token arithmetic exactness on specific datasets matters above all else, or if you require multimodal vision fine-tuning (vision behavior is inherited from the base and was not extensively evaluated here).
Practical notes: follow recommended sampling settings (temperature=0.6, top_p=0.95, top_k=20) and allow generous max_new_tokens (16k+) because outputs open with a learned deliberation span; the model inherits the teacher's tendency for over-long deliberation on easy problems. Weights are released under Apache-2.0 for research and experimentation.
