Two “missing” episodes of Doctor Who are coming to BBC iPlayer in the U.K., so if you have an account, you can stream them, save 73% and claim a voucher with NordVPN.
…

Chinese electronics and car manufacturer Xiaomi surprised the global AI community today with the release of MiMo-V2-Pro, a new 1-trillion parameter foundation model with benchmarks approaching those of U.S. AI giants OpenAI and Anthropic, but at around a seventh or sixth the cost when accessed over proprietary API — and importantly, sending less than 256,000 tokens-worth of information back and forth.

Led by Fuli Luo, a veteran of the disruptive DeepSeek R1 project, the release represents what Luo characterizes as a “quiet ambush” on the global frontier. Furthermore, Luo stated in an X post that the company does plan to open source a model variant from this latest release, ” when the models are stable enough to deserve it.”

By focusing on the “action space” of intelligence—moving from code generation to the autonomous operation of digital “claws”—Xiaomi is attempting to leapfrog the conversational paradigm entirely.

Prior to this foray into frontier AI, Beijing-based Xiaomi established itself as a titan of “The Internet of Things” and consumer hardware.

Globally recognized as the world’s third-largest smartphone manufacturer, Xiaomi spent the early 2020s executing a high-stakes entry into the automotive sector. Its electric vehicles (EVs), such as the SU7 and the recently launched YU7 SUV, have turned the company into a vertically integrated powerhouse capable of merging hardware, software, and now, advanced reasoning.

This pedigree in physical-world engineering informs MiMo-V2-Pro’s architecture; it is built to be the “brain” of complex systems, whether those systems are managing global supply chains or navigating the intricate scaffolds of an autonomous coding agent.

Technology: The architecture of agency

The central challenge of the “Agent Era” is maintaining high-fidelity reasoning over massive spans of data without incurring a prohibitive “intelligence tax” in latency or cost. MiMo-V2-Pro addresses this through a sparse architecture: while it houses 1T total parameters, only 42B are active during any single forward pass, making it roughly three times the size of its predecessor, MiMo-V2-Flash.

The model’s efficiency is rooted in an evolved Hybrid Attention mechanism. Standard transformers typically face a quadratic increase in compute requirements as context grows; MiMo-V2-Pro utilizes a 7:1 hybrid ratio (increased from 5:1 in the Flash version) to manage its massive 1M-token context window. This architectural choice allows the model to maintain a deep “memory” of long-running tasks without the performance degradation usually seen in frontier models.

The analogy: Think of the model not as a student reading a book page-by-page, but as an expert researcher in a vast library. The 7:1 ratio allows the model to “skim” 85% of the data for context while applying high-density attention to the 15% most relevant to the task at hand.

This is paired with a lightweight Multi-Token Prediction (MTP) layer, which allows the model to anticipate and generate multiple tokens simultaneously, drastically reducing the latency required for the “thinking” phases of agentic workflows. According to Luo, these structural decisions were made months in advance, specifically to provide a “structural advantage” for the unexpected speed at which the industry shifted toward agents.

Product and benchmarking: A third-party reality check

Xiaomi’s internal data paints a picture of a model that excels in “real-world” tasks over synthetic benchmarks. On GDPval-AA, a benchmark measuring performance on agentic real-world work tasks, MiMo-V2-Pro achieved an Elo of 1426, placing it ahead of major Chinese peers like GLM-5 (1406) and Kimi K2.5 (1283).

While it still trails Western “max effort” models like Claude Sonnet 4.6 (1633) in raw Elo, it represents the highest recorded performance for a Chinese-origin model in this category.

The third-party benchmarking organization Artificial Analysis verified these claims, placing MiMo-V2-Pro at #10 on its global Intelligence Index with a score of 49. This places it in the same tier as GPT-5.2 Codex and ahead of Grok 4.20 Beta. These results suggest that Xiaomi has successfully built a model capable of the high-level reasoning required for engineering and production tasks.

Key metrics from Artificial Analysis highlight a significant leap over the previous open-weights version, MiMo-V2-Flash (which scored 41):

• Hallucination rate: The Pro model reduced hallucination rates to 30%, a sharp improvement over the Flash model’s 48%.

• Omniscience index: It scored a +5, placing it ahead of GLM-5 (+2) and Kimi K2.5 (-8).

• Token efficiency: To run the entire Intelligence Index, MiMo-V2-Pro required only 77M output tokens, significantly less than GLM-5 (109M) or Kimi K2.5 (89M), indicating a more concise and efficient reasoning process.

Xiaomi’s own charts further emphasize its “General Agent” and “Coding Agent” capabilities. On ClawEval, a benchmark for agentic scaffolds, the model scored 61.5, approaching the performance of Claude Opus 4.6 (66.3) and significantly outpacing GPT-5.2 (50.0). In coding-specific environments like Terminal-Bench 2.0, it achieved an 86.7, suggesting high reliability when executing commands in a live terminal environment.

How enterprises should evaluate MiMo-V2-Pro for usage

For the personas outlined in contemporary AI organizations—from Infrastructure to Security—MiMo-V2-Pro represents a paradigm shift in the “Price-Quality” curve.

Infrastructure decision-makers will find MiMo-V2-Pro a compelling candidate for the Pareto frontier of intelligence vs. cost. Artificial Analysis reported that running their index cost only $348 for MiMo-V2-Pro, compared to $2,304 for GPT-5.2 and $2,486 for Claude Opus 4.6.

For organizations managing GPU clusters or procurement, the ability to access top-10 global intelligence at roughly 1/7th the cost of Western incumbents is a powerful incentive for production-scale testing.

Data decision-makers can leverage the 1M context window for RAG-ready architectures, allowing them to feed entire enterprise codebases or documentation sets into a single prompt without the fragmentation required by smaller context models.

A systems/orchestration decision-maker should evaluate MiMo-V2-Pro as a primary “brain” for multi-agent coordination. Because the model is optimized for OpenClaw and Claude Code, it can handle long-horizon planning and precise tool use without the constant human intervention that plagues earlier models.

Its high ranking in GDPval-AA suggests it is particularly well-suited for the workflow and orchestration layer needed to scale AI across the enterprise. It allows for the creation of systems that can move beyond simple automation into complex, multi-step problem solving.

However, security decision-makers must exercise caution. The very “agentic” nature that makes the model powerful—its ability to use terminals and manipulate files—increases the surface area for prompt injection and unauthorized model access.

While its low hallucination rate (30%) is a defensive boon, the lack of public weights (unlike the Flash version) means internal security teams cannot perform the deep “model-level” audits sometimes required for highly sensitive deployments. Any enterprise implementation must be accompanied by robust monitoring and auditability protocols.

Pricing, availability, and the path forward

Xiaomi has priced MiMo-V2-Pro to dominate the developer market. The pricing is tiered based on context usage, with competitive rates for caching to support high-frequency reasoning tasks.

• MiMo-V2-Pro (up to 256K): $1 per 1M input tokens and $3 per 1M output tokens

• MiMo-V2-Pro (256K-1M): $2 per 1M input tokens and $6 per 1M output tokens

• Cache read: $0.20 per 1M tokens for the lower tier and $0.40 for the higher tier

• Cache write: Temporarily free ($0)

Here’s how it stacks up to other leading frontier models around the world:

This aggressive positioning is designed to encourage the high-intensity application flows that define the next generation of software. The model is currently available via Xiaomi’s first-party API only, with no current support for image or multimodal input—a notable omission in an era of “Omni” models, though Xiaomi has teased a separate MiMo-V2-Omni for those needs.

The “Hunter Alpha” period on OpenRouter proved that the market has a high appetite for this specific blend of efficiency and reasoning. Fuli Luo’s philosophy—that research velocity is fueled by a “genuine love for the world you’re building for”—has resulted in a model that ranks 2nd in China and 8th worldwide on established intelligence indices.

Whether it remains a “quiet” ambush or becomes the foundation for a global realignment of AI power depends on how quickly developers adopt the “action space” over the “chat window”. For now, Xiaomi has moved the goalposts: the question is no longer just “can it talk?” but “can it act?”

In the last few years, Chinese AI startup MiniMax has become one of the most exciting in the crowded global AI marketplace, carving out a reputation for delivering frontier-level large language models (LLMs) with open source licenses and before that, high-quality AI video generation models (Hailuo).

The release of MiniMax M2.7 today — a new proprietary LLM designed to perform well powering AI agents and as the backend to third-party harnesses and tools like Claude Code, Kilo Code and OpenClaw — marks yet a new milestone: Rather than relying solely on human-led fine-tuning, MiniMax has leveraged M2.7 to build, monitor, and optimize its own reinforcement learning harnesses.

This move toward recursive self-improvement signals a shift in the industry: a future where the models we use are as much the architects of their progress as they are the products of human research. The model is categorized as a reasoning-only text model that delivers intelligence comparable to other leading systems while maintaining significantly higher cost efficiency.

However, with M2.7 being proprietary for now, it is a sign once again that Chinese AI startups — for much of the last year, the standard-bearers in the world of the open source AI frontier, making them appealing for enterprises globally due to low (or no) costs and customization — are shifting strategy and pursuing more proprietary frontier models like U.S. leaders like OpenAI, Google, and Anthropic have been doing for years.

MiniMax becomes the second Chinese startup to release a proprietary cutting-edge LLM in recent months following z.ai with its GLM-5 Turbo, and rumors that Alibaba’s Qwen team is also shifting to proprietary development in the wake of the departure of senior leadership and other researchers.

Technical achievement: The self-evolution loop

The defining characteristic of MiniMax M2.7 is its role in its own creation. According to company documentation, earlier versions of the model were used to build a research agent harness capable of managing data pipelines, training environments, and evaluation infrastructure.

By autonomously triggering log-reading, debugging, and metric analysis, M2.7 handled between 30 percent and 50 percent of its own development workflow.

This is not merely an automation of rote tasks; the model optimized its own programming performance by analyzing failure trajectories and planning code modifications over iterative loops of 100 rounds or more.

“We intentionally trained the model to be better at planning and at clarifying requirements with the user,” explained MiniMax Head of Engineering Skyler Miao on the social network X. “Next step is a more complex user simulator to push this even further.”

This capability extends to complex environments via the MLE Bench Lite, a series of machine learning competitions designed to test autonomous research skills.

In these trials, M2.7 achieved a medal rate of 66.6 percent, a performance level that ties with Google’s new Gemini 3.1 and approaches the current state-of-the-art benchmarks set by Anthropic’s Claude Opus 4.6.

The goal, according to MiniMax, is a transition toward full autonomy in model training and inference architecture without human involvement.

Performance evolution: MiniMax m2.7 vs. m2.5

When compared to its predecessor, M2.5, released in February 2026, the M2.7 model demonstrates significant gains in high-stakes software engineering and professional office tasks.

While M2.5 was celebrated for polyglot code mastery, M2.7 is designed for real-world engineering—tasks requiring causal reasoning within live production systems.

Key performance metrics include:

• Software engineering: M2.7 scored 56.22 percent on the SWE-Pro benchmark, matching the highest levels of global competitors like GPT-5.3-Codex.

• Professional office delivery: In document processing, M2.7 achieved an Elo score of 1495 on GDPval-AA, which the company claims is the highest among open-source-accessible models.

• Hallucination reduction: The model scores plus one on the AA-Omniscience Index, a massive leap from the negative 40 score held by M2.5.

• Hallucination rate: M2.7 achieves a hallucination rate of 34 percent, which is lower than the rates of 46 percent for Claude Sonnet 4.6 and 50 percent for Gemini 3.1 Pro Preview.

• System comprehension: On Terminal Bench 2, the model scored 57.0 percent, demonstrating a deep understanding of complex operational logic rather than simple code generation.

• Skill adherence: On the MM Claw evaluation, which tests 40 complex skills exceeding 2,000 tokens each, M2.7 maintained a 97 percent adherence rate, a substantial improvement over the M2.5 baseline.

• Intelligence parity: The model’s reasoning capabilities are considered equivalent to GLM-5, yet it uses 20 percent fewer output tokens to achieve similar results.

The model’s evolution is further evidenced by its score of 50 on the Artificial Analysis Intelligence Index, representing an 8-point improvement over its predecessor in just one month, and also taking the 8th place overall globally in terms of its overall intelligence across benchmarking tasks in various domains.

Not all independent, third-party benchmarks show improvement for M2.7 over M2.5: On BridgeBench, a set of tasks designed by agentic AI coding startup BridgeMind to test a model’s performance for “vibe coding,” or turning natural language into working code, M2.5 scored 12th place while M2.7 scored 19th place.

Access, pricing, and integration

MiniMax M2.7 is a proprietary model available through the MiniMax API and MiniMax Agent creation platforms. While the core model weights for M2.7 remain closed, the company continues to contribute to the ecosystem through the open-source interactive project OpenRoom.

For direct API integration and via third-party provider OpenRouter, MiniMax M2.7 maintains a cost-leading price point of 0.30 dollars per 1 million input tokens and 1.20 dollars per 1 million output tokens, which is unchanged from the pricing for M2.5.

To support different usage scales and modalities, MiniMax offers a structured Token Plan with various subscription tiers. These plans allow users to access models across text, speech, video, image, and music under a single unified quota.

To further drive adoption, MiniMax has launched an Invite and Earn referral program, providing a 10 percent discount to new invitees and a 10 percent rebate voucher to the inviter.

Monthly standard Token Plan pricing: The standard monthly tiers are designed for entry-level developers to heavy regular users.

• Starter: $10 per month for 1,500 requests per 5 hours.

• Plus: $20 per month for 4,500 requests per 5 hours.

• Max: $50 per month for 15,000 requests per 5 hours.

Monthly high-speed Token Plan pricing: For production-scale workloads requiring the M2.7-highspeed variant, the following tiers are available:

• Plus-Highspeed: $40 per month for 4,500 requests per 5 hours.

• Max-Highspeed: $80 per month for 15,000 requests per 5 hours.

• Ultra-High-Speed: $150 per month for 30,000 requests per 5 hours.

Yearly Token Plan pricing: Yearly subscriptions provide significant discounts for long-term commitment:

• Standard Starter: $100 per year (saves 20 dollars).

• Standard Plus: $200 per year (saves 40 dollars).

• Standard Max: $500 per year (saves 100 dollars).

• High-Speed Plus: $400 per year (saves 80 dollars).

• High-Speed Max: $800 per year (saves 160 dollars).

• High-Speed Ultra: $1,500 per year (saves 300 dollars).

One request in these plans is roughly equivalent to one call to MiniMax M2.7, though other models in the suite, such as video or high-definition speech, consume requests at a higher rate.

Official tool integrations

To ensure seamless adoption, MiniMax has provided official documentation for integrating M2.7 into over 11 major developer tools and agent harnesses.

This includes widely used platforms such as Claude Code, Cursor, Trae, and Zed. Other officially supported tools include OpenCode, Kilo Code, Cline, Roo Code, Droid, Grok CLI, and Codex CLI.

Additionally, the model supports the Model Context Protocol, allowing it to natively use tools like Web Search and Understand Image for multimodal reasoning. Developers using the Anthropic SDK can easily integrate M2.7 by modifying the ANTHROPIC_BASE_URL to point to the MiniMax endpoint.

When using MiniMax as a provider in tools like OpenClaw, image understanding capabilities are automatically configured via the model’s VLM API endpoint, requiring no extra setup from the user.

With its deep bench of integrations and its pioneering approach to recursive self-evolution, MiniMax M2.7 represents a significant step toward an AI-native future where models are as involved in their own progress as the humans who guide them.

Strategic implications for enterprise decision-makers

Technical decision-makers should interpret the M2.7 release as evidence that agentic AI has moved from theoretical prototyping to production-ready utility.

The model’s ability to reduce recovery time for live production incidents to under three minutes by autonomously correlating monitoring metrics with code repositories suggests a paradigm shift for SRE and DevOps teams.

Enterprises currently facing pressure to adopt AI-driven efficiencies must decide whether they are content with AI as a sophisticated assistant or if they are ready to integrate native agent teams capable of end-to-end full project delivery.

From a financial perspective, M2.7 represents a significant breakthrough in cost efficiency for high-level reasoning. Analysis indicates that M2.7 costs less than one-third as much to run as GLM-5 at equivalent intelligence levels.

For example, running a standard intelligence index cost 176 dollars on M2.7 compared to 547 dollars for GLM-5 and 371 dollars for Kimi K2.5. This aggressive pricing strategy places M2.7 on the Pareto frontier of the intelligence vs. cost chart, offering enterprise-level reasoning at a fraction of the market rate.

The current market is saturated with high-performance models, many of which still hold slight edges in general reasoning scores. But the specific optimization of M2.7 for Office Suite fidelity in Excel, PPT, and Word and its high performance in the GDPval-AA benchmark make it a primary candidate for organizations focused on professional document workflows and financial modeling.

Decision-makers must weigh the benefits of a general-purpose frontier model against a specialized engine like M2.7, which is built to interact with complex internal scaffolds and toolsets.

Ultimately, the fact that it is fielded by a Chinese company (headquartered in Shanghai) and subject to that country’s laws in addition to the user’s country, and is not available for offline or local usage yet, may make it a tough sell for enterprises operating in the U.S. and the West — especially those in highly-regulated or government-facing industries.

Nonetheless, the shift toward self-evolving models suggests that the ROI of AI investment will increasingly be tied to the recursive gains of the system itself.

Organizations that adopt models capable of improving their own harnesses may find themselves on a faster iteration curve than those relying on static, human-only refinement. With MiniMax’s aggressive integration into the modern developer stack, the barrier to testing these autonomous workflows has dropped significantly, placing pressure on competitors to deliver similar native agent capabilities.