Xiaomi has released MiMo V2.5 Pro, an open-weights model that has landed at the top of the Artificial Analysis Intelligence Index, tying with Moonshot’s Kimi K2.6 for first place among all publicly available open-weights models. The model was released on April 22, 2026, and its benchmark performance has drawn significant attention from the developer community since then.
MiMo V2.5 Pro scores 54 on the Artificial Analysis Intelligence Index, matching Kimi K2.6 and placing it ahead of every other open-weights model currently on the market. The model uses a mixture-of-experts architecture with 1 trillion total parameters, of which 42 billion are active during any given inference call. It ships with a 1-million-token context window and is designed to handle agentic task chains involving more than 1,000 sequential tool calls, a capability threshold that is increasingly relevant as developers build autonomous workflows rather than single-turn applications.
Benchmark Scores That Challenge Closed-Source Rivals
On SWE-bench Pro, the industry-standard coding benchmark, MiMo V2.5 Pro scores 57.2%. On the Humanity’s Last Exam benchmark, it reaches 34%, an improvement of 6 percentage points over its predecessor. On IFBench, which tests instruction-following accuracy, it scores 80%, up 11 percentage points from the previous version. These are not marginal improvements — they represent a consistent upward trajectory across the three benchmarks most commonly used to evaluate frontier model capabilities.
The predecessor, MiMo V2 Pro, was released on March 18, 2026. The gap between the two releases is just over five weeks, which is a notably fast iteration cycle for a model at this scale. The speed of iteration suggests that Xiaomi’s AI research team has developed a reliable pipeline for model development and evaluation, rather than treating each release as a one-off effort.
According to Startup Fortune, running the full Artificial Analysis benchmark suite on MiMo V2.5 Pro costs $462, compared to $948 for Kimi K2.6. That is a cost difference of more than 50% for equivalent benchmark coverage, which has direct implications for developers and enterprises evaluating which open-weights model to deploy in production.
Pricing and the Open-Weights Competitive Landscape
MiMo V2.5 Pro is priced at $1.00 per million input tokens and $3.00 per million output tokens. For context, Anthropic’s Claude Opus 4.5, the closed-source model that DeepSeek V4 Pro claims to approach on agentic tasks, is priced at $15 per million input tokens and $75 per million output tokens. The price differential is not a minor discount; it is an order-of-magnitude difference that fundamentally changes the economics of deploying frontier-class AI in production environments.
The open-weights model market has become one of the most competitive segments of the global AI industry, and Chinese labs are now its dominant force. Xiaomi’s MiMo V2.5 Pro, Moonshot’s Kimi K2.6, Alibaba’s Qwen family, and DeepSeek’s V4 Flash all compete for the same developer audience: teams that want frontier-class performance without the cost structure or vendor lock-in of closed-source APIs. The fact that MiMo V2.5 Pro ties Kimi K2.6 for first place on the Artificial Analysis Index means that two of the top three open-weights models in the world are now Chinese.
(Related: Moonshot AI’s Kimi K2.6 Code Preview Quietly Outperforms Claude Opus 4.5 at 76% Lower Cost)
The open-source weights for MiMo V2.5 Pro had not yet been publicly released as of April 25, 2026. Xiaomi has indicated that the weights are coming but has not specified a release date. The API is now available, so developers can evaluate the model’s capabilities before deciding whether to wait for the weights or build on the hosted API.
Xiaomi’s Broader AI Ambitions
Xiaomi’s entry into the frontier AI model market is relatively recent but has accelerated quickly. The company is better known globally for its consumer electronics, such as smartphones, smart home devices, and, more recently, electric vehicles but its AI research division has been building quietly and is now producing models that compete directly with dedicated AI labs. The MiMo family name stands for “Mixture of Models,” reflecting the MoE architecture that allows the model to activate only a fraction of its total parameters during inference, keeping costs low while maintaining high performance on complex tasks.
The timing of MiMo V2.5 Pro’s release one day before DeepSeek’s V4 full open-source release means that the week of April 21-26 has produced two Chinese open-weights models that each claim the top position on major benchmarks. For developers evaluating which foundation model to build on, the choice is no longer between a Chinese open-source model and a US closed-source model. It is increasingly a choice among multiple Chinese open-source models, each with different strengths, pricing structures, and deployment characteristics.
The Agentic Capability Threshold
One of the most commercially significant specifications of MiMo V2.5 Pro is its support for agentic task chains involving more than 1,000 sequential tool calls. This is a threshold that matters because most real-world agentic workflows, such as automated research pipelines, multi-step code generation and testing, and document processing chains, involve far more than the 10 to 20 tool calls that earlier models could reliably handle. The ability to sustain coherent reasoning and accurate tool use across hundreds of sequential steps is a prerequisite for the kind of autonomous workflows that enterprises are beginning to deploy at scale.
The 1-million-token context window compounds this advantage. Long-context capability and extended tool-call chains are complementary: a model that can process a million tokens in a single pass can maintain awareness of the full state of a complex task across many steps, rather than losing track of earlier context as the task progresses. For developers building agentic applications, the combination of these two capabilities in a model priced at $1/$3 per million tokens is a significant practical advantage over alternatives that offer only one capability or both at substantially higher cost.