tawer12/qwen3-4b-recursive-sft-v3.1

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 11, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The tawer12/qwen3-4b-recursive-sft-v3.1 is a 4 billion parameter, supervised-fine-tuned checkpoint based on Qwen3-4B-Base, specifically designed for recursive problem-solving. It was trained on component calls extracted from curated GSM8K teacher rollouts, incorporating special protocol tokens for recursive inference. This research checkpoint excels at mathematical reasoning by decomposing problems into sub-tasks, making it suitable for complex arithmetic and algebraic challenges rather than general-purpose chat.

Loading preview...

Overview

tawer12/qwen3-4b-recursive-sft-v3.1 is a 4 billion parameter model, a supervised-fine-tuned (SFT) checkpoint derived from Qwen3-4B-Base. It is specifically designed for recursive problem-solving, particularly in mathematical contexts. This model is a research checkpoint, not intended as a general-purpose chat assistant, and incorporates special protocol tokens for its recursive inference mechanism.

Key Capabilities

  • Recursive Problem Solving: Utilizes a unique recursive inference engine that breaks down complex problems into sub-tasks using actions like SOLVE_DIRECTLY, DECOMPOSE, and AGGREGATE.
  • Mathematical Reasoning: Fine-tuned on component calls from curated GSM8K teacher rollouts, making it proficient in arithmetic and algebraic problems.
  • Protocol-Driven Inference: Requires a specific RecursiveEngine and protocol tokens for execution, rather than standard chat templates or single model.generate calls.
  • Specialized Training: Trained on raw recursive prompt strings with an output-token-only causal language modeling objective, using 20,332 component-level SFT examples.

Good For

  • Mathematical Problem Decomposition: Ideal for tasks requiring the breakdown of complex math problems into manageable steps.
  • Research in Recursive AI: Useful for researchers exploring recursive reasoning architectures and protocol-based inference.
  • Specialized Mathematical Applications: Suitable for applications needing robust, step-by-step solutions to arithmetic and algebra problems, as demonstrated by historical scores of 70.68% on MATH algebra and 50.62% on MATH5000.

Finetunes
1 models