kaivoss/MiMo-V2.6-Distill-Qwen-9B-compat

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 22, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

MiMo-V2.6-Distill-Qwen-9B is a 9 billion parameter agentic model developed by Xiaomi MiMo, fine-tuned from Qwen3.5-9B. It specializes in agent tasks across coding, general automation, visual coding, and cybersecurity, demonstrating significant performance improvements over its base model. This model is designed as an SFT checkpoint for research into agentic reinforcement learning, offering enhanced capabilities in complex problem-solving. It supports a context length of 32768 tokens, making it suitable for detailed agentic workflows.

Loading preview...

MiMo-V2.6-Distill-Qwen-9B Overview

MiMo-V2.6-Distill-Qwen-9B is a 9 billion parameter agentic model developed by Xiaomi MiMo. It is a supervised fine-tuned (SFT) version of Qwen3.5-9B, trained on MiMo-generated data to excel in various agentic tasks. This model is released as a foundational checkpoint for open research in agentic reinforcement learning.

Key Capabilities & Performance

The model demonstrates substantial performance gains over its base model, Qwen3.5-9B, across multiple domains:

  • Code Generation: Significantly improved scores on SWE Verified (61.1 vs 60.0), SWE Pro (44.6 vs 32.0), and MiMo Code (51.6 vs 19.5).
  • Cybersecurity: Achieves 31.3 on MiMo Cyber compared to 5.7.
  • General Agent Tasks: Shows strong performance in AutomationBench (30.3 vs 5.0), Terminal Bench (37.1 vs 27.0), Toolathlon-Verified (35.2 vs 25.9), OfficeQA (19.5 vs 9.0), JobBench (18.3 vs 2.6), and MiMo General (62.2 vs 28.5).
  • Visual Coding: Improved to 64.0 on MiMo Visual Coding from 61.7.

These improvements stem from a diverse training data mixture totaling 77.4 billion tokens, with 27.2 billion loss-bearing tokens, distributed across code, cyber, general, and visual domains.

Use Cases

This model is particularly well-suited for:

  • Agentic Reinforcement Learning Research: Serves as an SFT starting point for developing and experimenting with agentic RL systems.
  • Automated Task Execution: Excels in general automation, terminal operations, and tool-use scenarios.
  • Code Development & Cybersecurity: Offers enhanced capabilities for code generation and cybersecurity-related tasks.
  • Visual Coding Applications: Can be applied to tasks involving visual programming or interpretation.

Finetunes
1 models