sriq-ai/sriq-mimo_v2.6-9b-v1.6

Hugging Face
VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 24, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

The sriq-ai/sriq-mimo_v2.6-9b-v1.6 is a 9.4 billion parameter language model developed by SRIQ, fine-tuned from XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B. This model is uniquely designed for efficient reasoning, performing its internal thought process in compressed Simplified Chinese regardless of the input language, and then providing the final answer in the user's language. It excels in mathematical and coding question-answering tasks, featuring a 32768 token context length and a self-correction mechanism during its reasoning trace.

Loading preview...

Overview

sriq-ai/sriq-mimo_v2.6-9b-v1.6 is a 9.4 billion parameter model developed by SRIQ, based on XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B (which is an SFT of Qwen/Qwen3.5-9B). It features a unique reasoning mechanism where it processes its internal thoughts in compressed Simplified Chinese, then provides the final answer in the user's requested language. This approach aims to reduce token cost for reasoning steps while maintaining answer quality.

Key Capabilities

  • Multilingual Reasoning: Performs internal reasoning in Simplified Chinese, regardless of the input language, and outputs answers in the user's language.
  • Self-Correction: The v1.6 dataset introduces a self-correction mechanism, where 19.5% of training rows include a planted mistake that the model's 验： (check) block identifies and redoes.
  • Efficient Trace: Chinese reasoning uses roughly half the tokens English needs, making the reasoning trace more cost-effective.
  • Mathematical & Coding Focus: Fine-tuned on a dataset of 9,988 conversations covering math and coding question/answer pairs.
  • High Context Length: Supports a context length of 262,144 tokens, inherited from its base architecture.

When to Use This Model

  • Cost-Effective Reasoning: Ideal for applications where complex reasoning is required, and optimizing token usage for internal thought processes is beneficial.
  • Mathematical Problem Solving: Suited for tasks involving mathematical questions and problem-solving, leveraging its specialized training.
  • Coding Assistance: Can be used for coding-related question answering, building on its base as an agentic coding model.
  • Multilingual Applications: Useful in scenarios where users interact in various languages but benefit from a consistent, efficient internal reasoning process.

Limitations

  • No benchmarks were run for this release, so performance metrics against the base model are not available.
  • The reasoning trace will always be in Chinese if displayed to users.
  • Tool-use behavior is untested after fine-tuning, as the training data did not include tool calls.

Finetunes
2 models