'Better than DeepSeek': Xiaomi's MiMo-V2.6-Pro debuts as the top open weights model in the world

Xiaomi has officially entered the high-stakes arena of frontier-class open-weights modeling with the release of MiMo-V2.6-Pro. According to reporting from VentureBeat (https://venturebeat.com/technology/better-than-deepseek-xiaomis-mimo-v2-6-pro-debuts-as-the-top-open-weights-model-in-the-world-alongside-cheaper-v2-6-flash), the new model has claimed the top spot on the Artificial Analysis Intelligence Index, outperforming several established proprietary and open-weight competitors.

What happened

The release of MiMo-V2.6-Pro marks a significant milestone for Xiaomi, as the model has achieved an Intelligence Index score of 46. This performance metric places it ahead of xAI's Grok 4.6, which holds a score of 44, and Google's Gemini 3.8 Flash, which sits at 41. The model also surpasses the performance of DeepSeek's recent offerings, including DeepSeek V4.1 Flash (39) and DeepSeek V4.1 Pro (36).

Alongside the flagship Pro model, Xiaomi has introduced MiMo-V2.6-Flash. This secondary model is designed for high-efficiency, low-cost applications. The pricing structure for the Flash variant is notably aggressive, set at $0.14 per million input tokens and $0.28 per million output tokens. For developers requiring higher reasoning capabilities, the Pro model is priced at $0.435 per million uncached input tokens and $0.87 USD per million output tokens.

The technical architecture of the Pro model is built around a massive 1-million-token context window. Furthermore, the model is natively multimodal, capable of processing text, images, audio, and video inputs. This breadth of input capability suggests that Xiaomi is targeting a future where LLMs act as central orchestrators for diverse data streams.

Why it matters

The arrival of a model that claims to outperform DeepSeek and major US-based proprietary models suggests a rapid convergence in the capabilities of open-weights models. Xiaomi's approach is not just about the weights themselves; the company is providing a comprehensive ecosystem for researchers. This includes the release of the technical report, over 7,000 reinforcement learning (RL) task environments, and an end-to-end RL framework.

A key technical differentiator reported by VentureBeat is Xiaomi's "You Only RL Once" strategy. This methodology involves a complex training regimen where different domains and agent harnesses are integrated into a single, large-scale reinforcement learning training run. The distribution of the training budget for the Pro model highlights the intensity of this process: 43.5% of the budget was dedicated to the training itself, 43.8% to producing rollouts, and 12.7% to the grading process.

This level of transparency and the release of the underlying RL environments could significantly accelerate the development of agentic workflows and specialized fine-tuning within the global AI research community. By releasing the task environments, Xiaomi is essentially providing the "gym" in which future models can be trained, potentially setting the standard for how RL-based training is benchmarked and executed.

Furthermore, the economic implications of the Flash model cannot be overstated. With input costs as low as $0.14 per million tokens, Xiaomi is positioning itself to capture the massive market of high-volume, low-complexity tasks, such as summarization, classification, and basic data extraction, which are the building blocks of the burgeoning agentic economy.

What to watch

The industry will be watching the real-world adoption and latency of these models in production environments. While the Pro model boasts an impressive output speed of 134 tokens per second, the true test will be how it maintains this performance under heavy, multi-modal workloads. High-speed inference is critical for real-time applications like voice assistants and autonomous agents, and Xiaomi's ability to deliver this at scale will be a key metric for enterprise adoption.

Furthermore, the competitive landscape for open-weights models is shifting. With Xiaomi's MiMo series now leading the Intelligence Index, the pressure is on other major players like Meta, xAI, and Google to respond with updated weights or more efficient architectures. The "You Only RL Once" strategy may also become a blueprint for other developers looking to optimize the massive computational costs associated with large-scale reinforcement learning.

As the cost of intelligence continues to drop, particularly with the highly economical Flash variant, we may see a surge in complex, agent-driven applications that were previously cost-prohibitive. The industry must also watch for how proprietary providers like OpenAI and Anthropic respond to this "open-weights" pressure; if the gap between proprietary and open models continues to shrink, the pricing power of closed-source providers may face significant downward pressure. The democratization of high-performance reasoning through Xiaomi's release is a clear signal that the era of "intelligence as a luxury" is rapidly coming to an end.