Voltar às notícias
ProdutoInstruções AI Understanding

Xiaomi MiMo-V2.6-Pro tops open-weights AI rankings

Xiaomi's new flagship model achieves the highest open-weights score on Artificial Analysis, featuring 1.02 trillion parameters and live-streamed reinforcement learning training.

4 min readRead the linked source
Source-provided image accompanying Xiaomi MiMo-V2.6-Pro tops open-weights AI rankings
Referência de fonteFonte registrada
Editora
unite.ai
Link da fonte
unite.aihttps://www.unite.ai/xiaomis-new-flagship-model-leads-open-weight-rankings-with-a-score-of-46/
Tipo de fonte
Fonte vinculada — o status da fonte primária não foi estabelecido.
Também citado

História revisada pela última vez

ContextoEntenda isso em 60 segundos

Comece aqui

Termos-chave

API (Interface de Programação de Aplicativo)
Uma maneira estruturada de um sistema de software enviar solicitações e receber respostas de outro sistema.
Aprendizagem por Reforço
Treinamento por sinais de recompensa onde um agente aprende ações que maximizam o retorno a longo prazo.
Cartão Modelo
Documentação que descreve o uso pretendido, métricas, limitações e riscos de um modelo.
Teste você mesmoQuestionário explicado sobre modelos de IA

O que mudou desde a publicação

  1. Publicado pela primeira vez
  2. This source provides detailed technical specifications, benchmark scores, and training methodology for the MiMo-V2.6-Pro release, including live-streamed RL data and open-sourced code, which expands on the initial announcement of the model's release.

O que aconteceu

Xiaomi released the MiMo-V2.6 series, with the flagship Pro model scoring 46 on Artificial Analysis' Intelligence Index, the highest for open-weights models. The release includes detailed technical specifications, open-sourced weights, and live-streamed RL training data.

Xiaomi announced the release of the MiMo-V2.6 series on September 22, 2026, led by the MiMo-V2.6-Pro flagship model. According to Unite.AI, the model achieved a score of 46 on Artificial Analysis' Intelligence Index v4.3.2, making it the highest-scoring open-weights model on that leaderboard. This score ties with the proprietary Grok 4.7 (xhigh) model and surpasses other open-weights competitors like Z AI’s GLM-5.3 and Kimi K3.

The MiMo-V2.6-Pro is a sparse mixture-of-experts model with 1.02 trillion total parameters and 42 billion activated per token. It supports a 1M-token context length and handles text, image, video, and audio inputs. The details a 70-layer backbone with 384 routed experts, where eight are active per token. Xiaomi open-sourced the weights on Hugging Face and ModelScope, along with the full technical report, training environments, and RL code under an MIT license.

A distinctive aspect of this release is the live-streaming of the production reinforcement-learning run. Xiaomi reported that the Pro and Flash models completed 30 RL steps over roughly 750,000 trajectories in under six days. The reported costs were approximately $2.62 million for Pro and $0.85 million for Flash. During this process, the average pass rate on training tasks rose by 12% for Pro and 25% for Flash, with significant improvements on the DeepSWE v1.1 benchmark.

The series includes three variants: MiMo-V2.6-Pro, MiMo-V2.6-Flash, and MiMo-V2.6-Pro-UltraSpeed. The UltraSpeed variant offers up to 20x faster output at the same quality. API pricing remains unchanged from the V2.5 series, with Pro costing $0.435 per million tokens for cache-miss input and $0.87 for output. The models are available via AI Studio, MiMo Code, MiMo Desktop, and the MiMo API Platform.

Detalhes da fonte: unite.ai

Por que isso importa

This release marks a significant milestone for open-weights AI, as Xiaomi's model now matches proprietary leaders in benchmark performance while offering transparent training methodologies. The live-streamed RL process and open-sourced code allow for unprecedented verification of training efficiency and cost, potentially shifting industry standards for reproducibility and competitive benchmarking in the open-source sector.

The MiMo-V2.6-Pro's performance places it at the forefront of open-weights AI, challenging the dominance of proprietary models in high-level reasoning and coding tasks. By achieving a score of 46 on the Intelligence Index, Xiaomi demonstrates that open-source models can now compete with top-tier proprietary systems like Grok 4.7 and GPT 6 Astra in specific benchmark categories.

The transparency of the training process is a major development for the AI industry. By live-streaming the RL run and open-sourcing the code, Xiaomi allows researchers and developers to verify the reported efficiency and cost metrics. This level of reproducibility is rare and could set a new standard for how AI companies validate their training methodologies and performance claims.

The cost-efficiency of the MiMo-V2.6-Pro is a significant practical implication for enterprises. With API pricing that remains competitive and a high performance-to-cost ratio, the model offers a viable alternative for organizations looking to deploy advanced AI capabilities without the high costs associated with proprietary APIs. The MIT license further facilitates commercial use and integration into existing workflows.

Interactive Mechanism

Mecanismo interativo: como realmente funciona

Explore a tecnologia subjacente a este desenvolvimento de forma interativa.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
Verificação de conceito interativo+10 Points
AI Models Explained Quiz

What is the best response when AI Models Explained makes a mistake in production?

O que assistir a seguir

Monitor independent verification of the reported RL training costs and pass-rate improvements, as well as the adoption of the MiMo-V2.6-Pro model in enterprise environments where cost-efficiency and open licensing are critical factors.

Independent verification of the reported RL training costs and performance improvements will be crucial. While Xiaomi provided detailed metrics, third-party audits or reproductions of the training process will help confirm the accuracy of these claims and assess the model's real-world performance beyond the reported benchmarks.

The adoption of the MiMo-V2.6-Pro in enterprise and research settings will indicate its practical utility. Monitoring how developers and companies integrate the model into their applications, particularly in areas like software engineering and scientific research, will provide insights into its broader impact and potential limitations.

The response from other AI companies to Xiaomi's release will be important to watch. Competitors may accelerate their own open-weights releases or adjust their pricing and performance strategies to maintain their market position. The competitive dynamics in the open-source AI space are likely to intensify in the coming months.

Guias e questionários relacionados

Modelos de IA explicadosTreinamento de IAO que é IA?Teste o que você sabe – experimente um teste gratuito de IAProcure um termo de IA em nosso glossário

Atualizações e correções

Esta história canônica é atualizada quando o evento em desenvolvimento muda materialmente. Seu URL e a data de publicação original nunca mudam.

  • This source provides detailed technical specifications, benchmark scores, and training methodology for the MiMo-V2.6-Pro release, including live-streamed RL data and open-sourced code, which expands on the initial announcement of the model's release.
Veja o registro de correções públicas
Achou isso útil?