Pluto-AI-Labs/Atlas-Frontier-Distill-3B

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 2, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Atlas-Frontier-Distill-3B is a 3.1 billion parameter coding model developed by Pluto AI Research Lab, distilled from frontier model traces (Kimi-K3, GPT-5.6, Fable-5) and built upon Qwen2.5-Coder-3B-Instruct. It specializes in coding and debugging tasks, demonstrating a 65.0% Pass@1 accuracy on HumanEval, an improvement over its base model. This model is optimized for efficient local and edge deployment, focusing on high-quality coding solutions and reasoning traces.

Loading preview...

Atlas-Frontier-Distill-3B: A Specialized Coding Model

Pluto AI Research Lab developed Atlas-Frontier-Distill-3B, an experimental 3.1 billion parameter model built on Qwen2.5-Coder-3B-Instruct. Its core purpose is to transfer the advanced coding and debugging capabilities of large frontier models into a compact, efficient model suitable for local and edge deployment.

Key Distillation & Training Details

  • Distillation Source: Traces from frontier models like Kimi-K3, GPT-5.6, and Fable-5.
  • Dataset: Trained using QLoRA on 15,746 high-quality coding conversations, aggressively filtered from over 20,375 low-quality samples to ensure learning of productive coding behavior.
  • Training Method: QLoRA (NF4 4-bit) with a sequence length of 1024 for 1 epoch.

Performance & Behavioral Improvements

  • HumanEval Pass@1: Achieves 65.0% on a subset of 20 complex HumanEval problems, an improvement over the base model's 60.0%.
  • Behavioral Delta: Analysis using llm-diff shows maintained instruction fidelity and reasoning consistency, with improved response style that reduces conversational bloat, leading to more direct code output.

Intended Use Cases

  • Local coding assistants and IDE integration.
  • Autonomous debugging agents.
  • Python code generation.
  • Deployment in GGUF and Ollama for edge inference.

Limitations

  • Optimized specifically for coding and debugging; may perform less effectively on open-domain conversation or creative writing tasks.
  • Distillation quality is directly dependent on the quality of the teacher traces.

Finetunes
1 models