# Model Release ## 2026-09-22 MiMo-V2.6 Series Release - **mimo-v2.6-pro：** Our most powerful flagship reasoning model — omni-modal, ultra-high performance, trillion-parameter — built for complex projects, long-horizon tasks, high-stakes work, cybersecurity, and research. - **mimo-v2.6-flash：** Full-modality, high-intelligence, low-cost reasoning model — the best balance for high-frequency calls and large-scale tasks in professional workflows. - **mimo-v2.6-pro-ultraspeed：** Flagship V2.6-Pro performance, up to 20x faster. Built for real-time and latency-sensitive workloads. ## 2026-06-02 mimo-v2.5-asr Released {/* feishu-style:text-align:left */} Model Introduction: - **Bilingual & Dialects:** Supports Chinese, English, code-switching, and various regional dialects (Wu, Cantonese, Minnan, Sichuanese). - **Lyrics Transcription:** High-accuracy Chinese/English lyrics transcription in mixed vocal-instrumental tracks. - **Robust in Complex Audio:** Excels in challenging environments (high noise, far-field, multi-speaker). - **Knowledge-Intensive AI:** Pinpoint accuracy for classical poetry, jargon, and proper nouns, with auto-punctuation. ## 2026-04-23 mimo-v2.5-pro Released {/* feishu-style:text-align:left */} Model Introduction: - **Trillion parameters, efficient architecture:** 1T total parameters | 42B activations | 1M ultra-long context - **Ultimate Agent Performance:** In high-intensity agent scenarios, it performs comparably to Claude Opus4.6 ## 2026-04-23 mimo-v2.5 Released {/* feishu-style:text-align:left */} Model Introduction: - **Native full-modal perception + 1M context:** Supports native understanding of images, videos, audio, and text, enabling cross-modal precise perception and long-range reasoning, with comprehensive perception capabilities ranking among the industry's forefront - **Powerful full-modal Agent capabilities:** It has native Agent execution capabilities, enabling it to efficiently complete complex tasks such as browsing, understanding, reasoning, and operation, with its performance in daily tasks comparable to that of **mimo-v2.5-pro** - **Combining Performance and Efficiency:** While maintaining leading capabilities, achieving superior token efficiency, and positioned at the Pareto frontier of performance and efficiency ## 2026-04-23 MiMo-V2.5-TTS Series Release {/* feishu-style:text-align:left */} Model Introduction: - **Premium Voice TTS:** Built-in with multiple high-quality premium voices, it has strong capabilities in understanding and adhering to style instructions, supports fine-grained control over speech rate, emotion, tone, etc., and meets the expression needs of multiple scenarios - **Timbre Design:** Supports quickly defining and generating new timbres through a single sentence, making timbre creation more intuitive and efficient - **Timbre Cloning:** Based on a small number of audio samples, it can reproduce the target timbre with high fidelity, while maintaining the consistency of timbre characteristics and possessing good generalization and stability ## 2026-03-18 mimo-v2-pro Release {/* feishu-style:text-align:left */} **Model overview:** - Uses hybrid architecture with a 1:7 ratio of Global Attention to Sliding Window Attention (SWA); - 1T total parameters, with 42B active parameters; - Supports an ultra-long context window of 1M tokens. {/* feishu-style:text-align:left */} **Model details:** https://platform.xiaomimimo.com/#/docs/news/v2-pro-release ## 2026-03-18 mimo-v2-omni Release {/* feishu-style:text-align:left */} **Model overview:** - Supports up to 256K context length; - Supports text, vision, and speech modalities. {/* feishu-style:text-align:left */} **Model details:** https://platform.xiaomimimo.com/#/docs/news/v2-omni-release ## 2026-03-18 mimo-v2-tts Release {/* feishu-style:text-align:left */} **Model overview:** - Pretrained on over 100 million hours of data, using a self-developed multi-codebook speech modeling architecture; - Offers unique capabilities such as style control, singing, and voice cloning. {/* feishu-style:text-align:left */} **Pricing:** free for a limited time. {/* feishu-style:text-align:left */} **Model details:** https://platform.xiaomimimo.com/#/docs/news/v2-tts-release ## 2026-02-04 mimo-v2-flash Update 1. **Upgraded Coding Capabilities in Thinking Mode:** Specifically optimized for programming scenarios, the Thinking Mode now achieves a score of **78.6** on SWE-Bench Verified. Both the resolution rate and the quality of code generation have been significantly improved. 1. **Substantial Boost in Tool Calling Accuracy:** Stability issues regarding tool usage have been resolved. Tool calling accuracy in Thinking Mode has surged from 64% to **97.0%** , greatly enhancing execution reliability in Agent scenarios. 1. **Enhanced Instruction Following & Reduced Hallucinations:** - **Instruction Following:** Improved adherence to specific instructions, achieving an **AA-IFBench score of 72**. - **Factuality:** Enhanced rigor in factual responses, with the **Non-Hallucination Rate updated to 52%** . 1. **Optimized Handling of Complex Tasks:** Performance on Arena-Hard (Hard Prompts) in Thinking Mode has been strengthened, with the score rising to **60.6**. The model now demonstrates superior performance when handling high-difficulty logic problems. 1. **More Efficient Chain-of-Thought (CoT):** By optimizing CoT generation strategies, the consumption of redundant tokens has been significantly reduced. In benchmarks such as AIME25 and HMMT, the average generation length has decreased by **13% to 30%** . This effectively lowers latency and token costs while maintaining model performance.
**mimo-v2-flash-0204** **mimo-v2-flash-0112** **mimo-v2-flash**
**SWE-Bench Verified**
**Non-Thinking**		 **73.7** 73.3 73.4
**SWE-Bench Verified**
**Thinking**		 **78.6** 74.2 -
**Arena-Hard(Hard Prompt)**
**Non-Thinking**		 **49.3** 52.7 46.0
**Arena-Hard(Creative Writing)**
**Non-Thinking**		 **85.0** 86.0 78.3
**Aren-Hard(Hard Prompt)**
**Thinking**		 **60.6** 58.3 54.1
**Arena-Hard(Creative Writing)**
**Thinking**		 **85.8** 90.4 86.2
**AA-IFBench** **72** - 64
**AA-Omniscience Accuracy** **19** - 27
**AA-Omniscience Non-Hallucination Rate** **52%** - 9%
**Tool call success rate**
**Thinking**		 **97.0%** 64% 44%
{/* feishu-style:text-align:left */}
**Benchmark** **mimo-v2-flash (Acc)** **mimo-v2-flash (Avg Tokens)** **mimo-v2-flash-0204 (Acc)** **mimo-v2-flash-0204 (Avg Tokens)** **Length Reduction Ratio (%)**
**AIME25** 94.8 26984 91.1 18879 **30.04%**
**HMMT_Feb_25** 94.2 29294 92.9 21470 **26.71%**
**LiveCodeBench-AA** 83.2 21488 84.9 18335 **14.67%**
**GPQA-Diamond** 83.7 15862 83.8 13659 **13.89%**
> Note: The model API call method and model name remain unchanged ## 2026-01-12 mimo-v2-flash Update 1. **Enhanced general capabilities:** Improved the model’s performance on a wide range of general-purpose tasks. 1. **Upgraded coding performance in Thinking mode:** Strengthened code generation quality in Thinking mode, especially for programming scenarios. 1. **Deep integration with Claude Code:** Fully supports using Thinking mode in Claude Code. - **Best practice:** Set Thinking as the default mode to achieve more stable, higher-quality code generation. 1. **Optimized Experience for Other Code Agents:** Synchronized improvements to the interaction experience and generation quality across code assistant tools (Code Scaffolds) such as Kilo, Cline, and Roo. 1. **Improved Stability & Instruction Following:** Enhanced output stability and significantly improved adherence to specific output formats.
**mimo-v2-flash-0112** **mimo-v2-flash**
**SWE-Bench Verified**
**Non-Thinking**		 **73.3** 73.4
**SWE-Bench Verified Thinking** **74.2** -
**Arena-Hard(Hard Prompt)**
**Non-Thinking**		 **52.7** 46.0
**Arena-Hard(Creative Writing)**
**Non-Thinking**		 **86.0** 78.3
**Arena-Hard(Hard Prompt)**
**Thinking**		 **58.3** 54.1
**Arena-Hard(Creative Writing)**
**Thinking**		 **90.4** 86.2
> Note: The model API call method and model name remain unchanged ## 2025-12-16 mimo-v2-flash Release {/* feishu-style:text-align:left */} **Model overview:** - Uses hybrid architecture with a 1:5 ratio of Global Attention to Sliding Window Attention (SWA), a window size of 128, native 32K context, and extended training up to 256K; - Introduces 3 MTP layers, delivering 2.5 to 3.7× faster inference. {/* feishu-style:text-align:left */} **Pricing:** input 0.1/M tokens, output 0.3/M tokens. {/* feishu-style:text-align:left */} **Model details:** [mimo-v2-flash: High-Efficiency Inference, Code & Agent Foundation Model](https://platform.xiaomimimo.com/#/docs/news/news20251216) {/* feishu-style:text-align:left */} **Usage guide:** [First API call](https://platform.xiaomimimo.com/#/docs/quick-start/first-api-call)