wesjos/Mimo-Qwen3.5-9B-Cold

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 28, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

wesjos/Mimo-Qwen3.5-9B-Cold is a 9.4 billion parameter language model built on the Qwen3.5 architecture, specifically fine-tuned for concise, decisive, and direct responses. This model utilizes Direct Preference Optimization (DPO) on 1,027 LIMA-style preference pairs, uniquely skipping Supervised Fine-Tuning (SFT) to achieve a "cold-style" alignment. It excels at delivering direct answers while preserving general intelligence and knowledge, making it suitable for applications requiring clear and unambiguous communication.

Loading preview...

Overview

wesjos/Mimo-Qwen3.5-9B-Cold is a 9.4 billion parameter language model based on the Qwen3.5 architecture, specifically designed for cold-style preference alignment. This model distinguishes itself by applying Direct Preference Optimization (DPO) directly to its base model, MiMo-V2.6-Distill-Qwen-9B, without prior Supervised Fine-Tuning (SFT). It was trained using 1,027 LIMA-style preference pairs to produce responses that are concise, decisive, and direct, while maintaining its inherent intelligence and general knowledge.

Key Capabilities

  • Concise and Direct Responses: Optimized for clear, unambiguous, and to-the-point communication.
  • Preference Alignment: Achieves a specific "cold-style" output through DPO, making it suitable for applications where brevity and decisiveness are paramount.
  • Preserves Base Model Intelligence: Despite the lightweight alignment, it retains the general knowledge and capabilities of its Qwen3.5-architecture base.
  • Multilingual Support: Supports both English and Chinese languages.
  • Extended Context Length: Features a native context length of 262,144 tokens.

Good for

  • Applications requiring brief and factual answers: Ideal for chatbots, summarization, or information retrieval systems where users expect direct responses.
  • Scenarios needing decisive output: Useful in contexts where ambiguity should be minimized.
  • Developers seeking a distinct response style: Offers a unique "cold-style" alignment for specific user experience requirements.

Limitations

  • Small Training Scale: The alignment is based on a relatively small dataset (1,027 pairs), primarily influencing style rather than introducing new capabilities or knowledge.
  • Potential for Higher Tool Hallucination: When used for function calling, it may exhibit a slightly higher rate of hallucination, requiring application-layer parameter schema validation.
  • Knowledge Cutoff: Inherits the knowledge cutoff of its base Qwen3.5 model, as no new knowledge was introduced during training.

Finetunes
2 models