tawer12/qwen3-4b-recursive-sft-v3.2

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 30, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The tawer12/qwen3-4b-recursive-sft-v3.2 is a 4 billion parameter Qwen3-based model, fine-tuned for recursive math-solving tasks with a 32768 token context length. It utilizes a unique recursive protocol for problem decomposition and aggregation, making it specialized for complex mathematical reasoning rather than general chat. This model is a research checkpoint focused on structured problem-solving through recursive inference.

Loading preview...

Overview

This model, tawer12/qwen3-4b-recursive-sft-v3.2, is a specialized 4-billion parameter Qwen3-based language model. It is a research math-solving model, specifically fine-tuned for recursive problem-solving rather than general-purpose chat. The model was trained from Qwen3-4B-Base with additional recursive protocol tokens and features a 32768 token context length.

Key Capabilities

  • Recursive Math-Solving: Designed to break down complex math problems into smaller, manageable steps using a recursive inference protocol.
  • Protocol-Driven Inference: Generates specific actions like SOLVE_DIRECTLY, DECOMPOSE, or AGGREGATE to guide the problem-solving process.
  • Structured Output: Outputs raw recursive prompt strings, not Qwen chat-template conversations, preserving special protocol tokens.
  • Dedicated Engine: Comes with a recursive_engine.py for standalone execution of complete problem-solving trees.

Training Details

The model was trained over 6 epochs on a corpus of 14,126 component-level examples, representing 5,241 unique source problems. It focuses on downsampled direct-solve examples while retaining decomposition and aggregation examples from an expanded V3.1 corpus. The training used a global batch size of 128 component calls and a maximum sequence length of 2,048 tokens.

Limitations

It's important to note that format validity and answer correctness are distinct. The model's decomposition paths may be dependent or underspecified, and aggregation receives only child answer summaries. No new benchmark accuracy claims are made with this V3.2 release.

Finetunes
1 models