Downloads · 30 days
183
38% of all-time downloads
Qwen3.8-max-Reasoning-Distilled
khazarai/Qwen3.8-max-Reasoning-Distilled
Qwen3.8-max-Reasoning-Distilled is a image-text-to-text model from khazarai. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as apache-2.0.
All-time downloads
476
Public
Parameters
873M
1.8 GB on disk
Likes
6
Public
Repo makeup
Click a slice to open those files.
.safetensors1.7 GB · 99%
Parameter types
How the weights are stored.
BF16873M · 100%
Try a prompt
Base models
Model card
From the Hugging Face model README
Qwen3.8-max-Reasoning-Distilled
- Benchmark: khazarai/Multi-Domain-Reasoning-Benchmark
- Total Questions: 100
| Model | Score |
|---|---|
| khazarai/Qwen3-4B-Qwen3.6-plus-Reasoning-Distilled | 48.18 |
| Qwen/Qwen3.5-0.8B | 45.82 |
Qwen3.8-max-Reasoning-Distilled is a compact, high-efficiency language model fine-tuned using Knowledge Distillation from Qwen3.8-max.
It specifically addresses the common issue of overthinking loops and repetitive hesitations found in smaller raw reasoning models. By learning directly from high-quality reasoning traces, this model delivers concise, structured, and decisive Chain-of-Thought (CoT) outputs without wasting tokens on circular verification.
Key Features
-
Streamlined Chain-of-Thought (CoT): Replaces verbose, rambling reasoning paths with direct, step-by-step logic.
-
Elimination of Overthinking: Suppresses repetitive inner monologue loops (e.g., "Wait, let me re-verify...") that consume unnecessary tokens.
-
Higher Inference Efficiency: Drastically reduces total generated token count while maintaining or improving final answer accuracy.
-
Distilled Logic from Qwen3.8-max: Captures the complex evaluation heuristics of the teacher model into a smaller, faster student architecture.
Reasoning Quality Comparison
Below is a comparison highlighting how distillation improves reasoning structure and eliminates repetitive loops:
| Feature | Qwen3.5-0.8B | Qwen3.8-max-Reasoning-Distilled |
|---|---|---|
| Logic Flow | Frequent self-doubt, circular checks, hesitation | Linear, structured, and hypothesis-driven |
| Token Economy | High token overhead spent on repetitive verification | Low token overhead with concise step evaluation |
| Option Analysis | Reiterates choices multiple times without deciding | Evaluates each option once with clear justification |
| Decision Speed | Slow convergence to final answer | Fast, decisive output generation |
Example Trace
Task: Identify the beneficial effect of carpooling from multiple choice options.
- Qwen3.5-0.8B: "Evaluate option D... wait, let me check option E... actually, let me re-evaluate D... wait, is there a trick? Let me double-check..." (Overthinking Loop)
- Qwen3.8-max-Reasoning-Distilled: "Systematically evaluates choices A through H in a single pass $\rightarrow$ identifies option D as the primary environmental benefit $\rightarrow$ concludes decisively." (Clean & Direct)
Limitations
While the model minimizes overthinking, extremely complex multi-step mathematical problems may still require prompting for explicit scratchpad steps. Always verify outputs when using the model for domain-critical tasks (e.g., medical, legal, or financial decisions).
- Task
- Image-Text-to-Text
- Library
- transformers
- Type
- qwen3_5
- License
- apache-2.0
- Languages
- en
- Created
- Aug 31, 2026
- Updated
- Sep 19, 2026