Merlin-Research/Pluto

Overview

Pluto is a 9 billion parameter coding and reasoning model from Merlin Research, based on the Qwen/Qwen3.5-9B-Base architecture. It is specifically engineered for precision and robustness in coding tasks, prioritizing error minimization over general fluency. A key differentiator is its 1,000,000 token context length, allowing it to process entire repositories or extensive conversation histories without needing chunking.

Key Capabilities

• Precision-First Design: Optimized to minimize errors, making it highly effective for critical coding tasks where correctness is paramount.
• Massive Context Window: Supports up to 1 million tokens, enabling coherent reasoning across large codebases and multi-file diffs.
• Agentic Deployment Ready: Fine-tuned for seamless integration with agentic coding environments like Claude Code, OpenAI Codex, and local large-codebase workflows.
• Quantum Entropy Regularization (AER): Utilizes quantum noise from IBM Quantum Kingston during RL training to enhance robustness, prevent entropy collapse, and improve stability on out-of-distribution inputs.
• Distillation from Frontier Models: Incorporates knowledge from advanced coding models and a private dataset of reasoning traces to transfer deep reasoning capabilities at a 9B parameter scale.

Good For

• Complex code generation and refactoring.
• Multi-file codebase analysis and code review.
• Deployment in agentic coding pipelines (e.g., Claude Code, OpenAI Assistants API).
• Architecture planning and technical reasoning tasks.
• Local deployment for large private codebases using GGUF/quantized variants.