Blackfrost-AI/MiMo-V2.6-Distill-Qwen-9B-Derisked-BF16

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 27, 2026License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Blackfrost-AI/MiMo-V2.6-Distill-Qwen-9B-Derisked-BF16 is a 9-billion parameter language model derived from XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B, which is based on Qwen/Qwen3.5-9B. This model was created using targeted directional weight modification (DWM) to reduce false-refusal friction, particularly for authorized security research and agentic workflows, while aiming to retain the parent model's general capabilities. It is a full standalone BF16 checkpoint, not an adapter or quantized build, and supports a 262,144-token maximum context setting.

Loading preview...

Overview

Blackfrost-AI/MiMo-V2.6-Distill-Qwen-9B-Derisked-BF16 is a 9-billion parameter language model, a BF16 research derivative of XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B. Its lineage traces back to Qwen/Qwen3.5-9B. This model was developed using targeted directional weight modification (DWM), rather than gradient training, to specifically reduce false-refusal friction. This makes it particularly suitable for authorized security research and agentic workflows, while aiming to preserve the general capabilities of its parent checkpoint.

Key Capabilities

  • Reduced False-Refusal: Engineered to minimize unwarranted refusals, enhancing utility for specific research and agentic tasks.
  • Full Standalone Checkpoint: Provided as a complete BF16 model, not an adapter or quantized version.
  • High Context Length: Inherits the parent's 262,144-token maximum context setting.
  • Tool Calling Support: Validated for native structured tool calling tests, including non-streaming, streaming, thinking-enabled, and tool-result round-trip requests, utilizing MiMo reasoning and qwen3_coder tool-call parsers.
  • Multimodal Ready: Includes multimodal processor and vision files, though multimodal behavior has not been re-benchmarked on this derivative.

Good For

  • Security Research: Ideal for authorized security research where reduced false-refusal is beneficial.
  • Agentic Workflows: Suited for developing and deploying AI agents that require less friction in their responses.
  • Development & Experimentation: A robust base for further research and development, particularly for exploring DWM effects.
  • Qwen Ecosystem Users: Integrates well with existing Qwen-based workflows and tools, maintaining the original tokenizer and chat template.

Finetunes
1 models