Tuwhy/Qwen3-4B-OPSA

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 24, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Tuwhy/Qwen3-4B-OPSA is a 4 billion parameter Qwen3 causal language model enhanced with On-Policy Self-Adaptation (OPSA) for improved reasoning capabilities. Developed by Tuwhy, this model excels in mathematical and logical reasoning tasks, demonstrating significant performance gains over its base Qwen3-4B counterpart. It supports a native context length of 32,768 tokens, extendable to 131,072 tokens with YaRN, and features dynamic switching between thinking and non-thinking modes for varied use cases.

Loading preview...

Tuwhy/Qwen3-4B-OPSA: Enhanced Reasoning with On-Policy Self-Adaptation

This model is a 4 billion parameter Qwen3 checkpoint, developed by Tuwhy, that has been significantly improved through On-Policy Self-Adaptation (OPSA). OPSA is a novel training method that refines a policy without relying on a teacher model, reward model, or external task rewards, focusing on self-improvement by adapting to valid response tokens with low actor log probabilities.

Key Capabilities & Differentiators

  • Superior Reasoning: The OPSA-trained Qwen3-4B demonstrates substantial improvements in mathematical and logical reasoning. For instance, it achieves 62.08 / 83.33 on AIME24 and 58.44 / 83.33 on AIME25, significantly outperforming the base Qwen3-4B model.
  • Flexible Thinking Modes: Inherits Qwen3's unique ability to seamlessly switch between a thinking mode (for complex logical reasoning, math, and coding) and a non-thinking mode (for efficient, general-purpose dialogue). This allows for optimal performance across diverse scenarios.
  • Extended Context Window: Natively supports a context length of 32,768 tokens, which can be extended up to 131,072 tokens using the YaRN method for processing long texts.
  • Agentic Capabilities: Excels in tool-calling, integrating with frameworks like Qwen-Agent for precise interaction with external tools in both thinking and non-thinking modes.
  • Multilingual Support: Supports over 100 languages and dialects with strong multilingual instruction following and translation capabilities.

Ideal Use Cases

  • Complex Problem Solving: Excellent for applications requiring advanced mathematical, coding, or logical reasoning, especially where detailed thought processes are beneficial.
  • Dynamic Interaction: Suitable for conversational AI that needs to adapt its reasoning depth based on user input or task complexity.
  • Long Context Processing: Beneficial for tasks involving extensive documents or multi-turn conversations that exceed standard context limits.
  • Agent-based Systems: Highly effective for integrating with external tools and building sophisticated AI agents.

Finetunes
1 models