thegovind/blink-mimo-9b

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 24, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

The thegovind/blink-mimo-9b is a 9.41 billion parameter multimodal model developed by thegovind, built upon the XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B base. This model specializes in visual decision-making, specifically picking the next click from a screenshot by scoring offered targets as a typed choice, returning probabilities instead of generating text. It integrates its own vision encoder and excels at computer use tasks, processing both image tokens and text/JSON state for decision-making.

Loading preview...

Overview

thegovind/blink-mimo-9b is a 9.41 billion parameter multimodal model, based on the XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B architecture. It is designed for visual decision-making, particularly for tasks involving screen interaction. Unlike traditional LLMs that generate text, blink-mimo-9b processes visual input (screenshots) and text/JSON state to score potential actions, returning probabilities for typed choices rather than generating free-form text.

Key Capabilities

  • Multimodal Input: Accepts both image tokens from screenshots (via its integrated MiMo vision encoder) and text/JSON state for decision-making.
  • Probabilistic Output: Provides probabilities for a set of predefined choices (e.g., which button to click next), rather than generating conversational text.
  • Computer Use Automation: Optimized for tasks like identifying the next click on a screen, demonstrated by completing 133/230 (58%) zero-shot tasks in a computer use benchmark.
  • Fine-tuning Support: Allows for customization with LoRA adapters to fine-tune its text-side decision-making on specific datasets.
  • Efficient Readout: Utilizes a single forward pass to score options, applying FP32 softmax on next-token logits for offered labels.

Good for

  • Automated UI Interaction: Ideal for agents that need to make precise decisions on graphical user interfaces based on visual and contextual information.
  • Decision-making Systems: Suitable for applications requiring a probabilistic choice from a set of options, rather than open-ended text generation.
  • Research in Multimodal Agents: Provides a foundation for exploring and developing agents that interact with digital environments through visual perception and structured decision logic.

Limitations

  • Primarily English-centric.
  • Does not chat or explain its answers.
  • Weak spots include long policies and complex date/number reasoning.

Finetunes
1 models