nanoswe/qwen-3.6-35b-a3b-teacher

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 5, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

nanoswe/qwen-3.6-35b-a3b-teacher is a 35.1 billion parameter language model, a lightly fine-tuned variant of Qwen/Qwen3.6-35B-A3B, developed by nanoswe. It is specifically trained to generate agentic software-engineering trajectories, serving as a teacher model for distillation experiments. This model excels at producing SWE-bench-style trajectory data for training smaller student models, focusing on correct, concise, and non-cheating solutions. Its primary use is for generating high-quality software engineering problem-solving demonstrations.

Loading preview...

Model Overview

nanoswe/qwen-3.6-35b-a3b-teacher is a 35.1 billion parameter model derived from Qwen/Qwen3.6-35B-A3B. Its core purpose is to act as a "teacher" model, generating high-quality, agentic software-engineering trajectories. These trajectories are then used to train smaller "student" models through distillation.

Key Characteristics and Training

This model underwent a light fine-tuning process (one epoch at a learning rate of 6e-6) to sharpen its ability to produce structured trajectories and improve success rates on software engineering tasks, without significantly altering its base distribution. The training data consists of 14,268 curated trajectories generated by the base Qwen3.6-35B-A3B model on SWE-smith task instances. The curation process ensured:

  • Correctness: Only trajectories that successfully resolved their instances.
  • Integrity: Exclusion of trajectories matching known "cheat-patterns."
  • Difficulty: Removal of instances solved too easily (over 75% success rate).
  • Efficiency: Preference for the top-3 shortest correct trajectories per instance.
  • Context Length: All trajectories were within the 32,768 token limit.

The model was trained using full-parameter bf16 SFT with a sequence length of 32,768 and a batch size of 64 sequences.

Usage Notes

When using this model, it is crucial to:

  • Employ the bundled chat_template.jinja for plain <|im_start|>role\ncontent<|im_end|> turns.
  • Ensure enable_thinking=False, as the model was fine-tuned on trajectories generated without internal thinking steps.

Finetunes
1 models