Chinese electronics and car maker Xiaomi today surprised the global AI community with the release of MiMo-V2-Pro, a new 1-trillion parameter foundation model, whose benchmarks are close to those of American AI giants OpenAI and Anthropic, but the cost is about a seventh or sixth when accessed over proprietary APIs – and crucially, sending less than 256,000 tokens-worth of information back and forth.
Led by Fuli Luo, veteran of the disruptive DeepSeek R1 project, this release features Luo. "silent ambush" On the global border. Additionally, Luo said in an X post that the company is planning to open source a model version from this latest release, " "
by focusing on "action space" Intelligence—Moving from Code Generation to Autonomous Operation of Digital "claws"-Xiaomi is attempting to leapfrog the conversation paradigm entirely.
Before this foray into frontier AI, Beijing-based Xiaomi had established itself as a giant "internet of things" and consumer hardware.
Globally recognized as the world’s third-largest smartphone maker, Xiaomi executed a high-risk entry into the automotive sector in early 2020. Its electric vehicles (EVs), such as the SU7 and the recently launched YU7 SUV, have transformed the company into a vertically integrated powerhouse capable of merging hardware, software and, now, advanced logic.
This lineage in physical-world engineering informs the architecture of the MiMo-V2-Pro; it is designed to be "Brain" Of complex systems, whether those systems are managing global supply chains or navigating the complex scaffolds of an autonomous coding agent.
"agent era" Maintaining high-fidelity logic over vast expanses of data without taking any prohibitive action "intelligence tax" In latency or cost. The MiMo-V2-Pro addresses this through a sparse architecture: while it has 1T total parameters, only 42B are active during any single forward pass, making it approximately three times the size of its predecessor, the MiMo-V2-Flash.
The efficiency of the model lies in a developed hybrid attention mechanism. Standard transformers typically face a quadratic increase in computation requirements as the reference increases; The MiMo-V2-Pro uses a 7:1 hybrid ratio (increased from 5:1 in the Flash version) to manage its huge 1M-token reference window. This architectural choice allows the model to maintain depth "Memory" Long running tasks without performance degradation are typically seen in Frontier models.
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, applying a high-density focus on 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 predict and generate multiple tokens simultaneously, significantly reducing the latency required for this. "Thinking" phases of agentic workflows. According to Luo, these structural decisions were made months ago specifically to provide "structural advantage" The unprecedented speed at which the industry shifted toward agents.
Products and Benchmarking: A Third-Party Reality Check
Xiaomi’s internal data paints a picture of a model that is excellent "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, putting it ahead of major Chinese counterparts such as GLM-5 (1406) and KMK2.5 (1283).
"maximum effort" Models like the Cloud Sonnet 4.6 (1633) in raw Aloe, this represents the highest recorded performance for a Chinese-origin model in this category.
Third-party benchmarking organization Artificial Analysis verified these claims, ranking the MiMo-V2-Pro at #10 on its Global Intelligence Index with a score of 49. This puts it at the same level as the GPT-5.2 codec and ahead of the Grok 4.20 beta. These results show that Xiaomi has successfully created a model capable of performing the high-level reasoning required for engineering and production tasks.
Key metrics from artificial intelligence analysis highlight a significant leap compared to the previous open-weighted version, MiMo-V2-Flash (which scored 41):
- Hallucination Rate: The Pro model reduced the hallucination rate by 30%, which is a sharp improvement from the Flash model’s 48%.
-
Omniscience Index: It scored +5, putting it ahead of GLM-5 (+2) and KM K2.5 (-8).
-
Token Efficiency: To run the entire intelligence index, MiMo-V2-Pro requires only 77M output tokens, which is significantly less than GLM-5 (109M) or KmK2.5 (89M), indicating a more concise and efficient reasoning process.
"general agent" And "coding agent" Abilities. On Cloudeval, a benchmark for agentic scaffolds, the model scored 61.5, coming close to the performance of Cloud Opus 4.6 (66.3) and well ahead of GPT-5.2 (50.0). In a coding-specific environment like Terminal Bench 2.0, it achieved 86.7, suggesting high reliability when executing commands in a live terminal environment.
How should enterprises evaluate MiMo-V2-Pro for use
For the aforementioned personalities in contemporary AI organizations – from infrastructure to security – MiMo-V2-Pro represents a paradigm shift "value quality" curve.
Those making decisions about infrastructure will find MiMo-V2-Pro an attractive candidate for the Pareto front of intelligence versus cost. Artificial Analysis reported that the cost of running their index for MiMo-V2-Pro is only $348, compared to $2,304 for GPT-5.2 and $2,486 for Cloud Opus 4.6.
For organizations managing GPU clusters or procurement, the ability to access top 10 global intelligence at approximately 1/7th the cost of Western incumbents is a powerful incentive for production-scale testing.
Data decision-makers can take advantage of the 1M context window for RAG-ready architectures, allowing them to feed the entire enterprise codebase or documentation set into a single prompt without the fragmentation required by smaller context models.
A system/orchestration decision-maker should evaluate the MiMo-V2-Pro as a primary "Brain" for multi-agent coordination. Because the model is optimized for OpenGL and cloud code, it can handle long-horizon planning and precise equipment usage without the constant human intervention that plagued earlier models.
Its high ranking in GDPval-AA shows that it is particularly well-suited for the workflow and orchestration layer needed to scale AI across the enterprise. This allows the creation of systems that can move beyond simple automation to complex, multi-step problem solving.
However, security decision makers should exercise caution. very much "agentic" The nature that makes the model powerful – its ability to use terminals and manipulate files – increases the surface area for quick injection and unauthorized model access.
While its low hallucination rate (30%) is a defensive boon, the lack of public loadout (unlike the Flash version) means that internal security teams cannot perform intensive work. "model level" Highly sensitive deployments sometimes require audits. Any enterprise implementation should be accompanied by robust monitoring and auditability protocols.
Xiaomi has priced the MiMo-V2-Pro to dominate the developer market. Pricing is based on context usage, with competitive rates for caching to support high-frequency logic functions.
- $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
-
Read Cache: $0.20 per 1M tokens for the lower tier and $0.40 for the higher tier
-
Write Cache: Temporarily Free ($0)
Here’s how it matches up with other leading edge models from around the world:
| Sample |
input |
Production |
total cost |
Source |
|
grok 4.1 fast |
$0.20 |
$0.50 |
$0.70 |
xai |
|
Minimax M2.7 |
$0.30 |
$1.20 |
$1.50 |
minimal maximum |
|
gemini 3 flash |
$0.50 |
$3.00 |
$3.50 |
|
|
KM-K2.5 |
$0.60 |
$3.00 |
$3.60 |
moon |
|
MiMo-V2-Pro(≤256K) |
$1.00 |
$3.00 |
$4.00 |
xiaomi mimo |
|
glm-5-turbo |
$0.96 |
$3.20 |
$4.16 |
openrouter |
|
GLM-5 |
$1.00 |
$3.20 |
$4.20 |
Z.ai |
|
cloud haiku 4.5 |
$1.00 |
$5.00 |
$6.00 |
anthropic |
|
quen3-max |
$1.20 |
$6.00 |
$7.20 |
alibaba cloud |
|
gemini 3 pro |
$2.00 |
$12.00 |
$14.00 |
|
|
GPT-5.2 |
$1.75 |
$14.00 |
$15.75 |
OpenAI |
|
GPT-5.4 |
$2.50 |
$15.00 |
$17.50 |
OpenAI |
|
cloud sonnet 4.5 |
$3.00 |
$15.00 |
$18.00 |
anthropic |
|
cloud opus 4.6 |
$5.00 |
$25.00 |
$30.00 |
anthropic |
|
GPT-5.4 Pro |
$30.00 |
$180.00 |
$210.00 |
OpenAI |
This aggressive positioning is designed to encourage the high-intensity application flows that define next-generation software. The model is currently only available through Xiaomi’s first-party API, with no existing support for image or multimodal input – a notable omission in an era "omni" model, although Xiaomi has teased a separate MiMo-V2-Omni for those needs.
"hunter alpha" The period on OpenRouter proved that there is a high appetite in the market for this specific blend of efficiency and logic. Fuli Luo’s philosophy that fuels research momentum ""-This has resulted in a model that ranks 2nd in China and 8th worldwide on established intelligence indices.
whether he remains one "Calm" AI could become an ambush or the basis for a global realignment of power depending on how quickly developers adopt it "action space" Above "chat window". For now, Xiaomi has moved the goalposts: the question is no longer justified "Can it talk?" But "Can this work?"
<a href