Xiaomi’s MiMo-V2.6-Pro Is Now the Strongest Open-Weight AI Model

Set Trending Topics as a preferred source on Google.

Chinese electronics group Xiaomi recently released its MiMo-V2.6 model family and published the weights on Hugging Face under an MIT license. The flagship, MiMo-V2.6-Pro, scores 46 points on the Artificial Analysis Intelligence Index, making it the highest-rated open-weight model available anywhere. It takes the crown from Z.ai’s GLM-5.3 and Moonshot AI’s Kimi K3, which both sit at 44 points.

The jump over Xiaomi’s own previous generation is steep: MiMo-V2.5-Pro reached 26 points on the same index. The family also includes the smaller MiMo-V2.6-Flash and MiMo-V2.6-Pro-UltraSpeed, a variant that delivers up to 20 times the output speed at what Xiaomi says is identical quality.

How It Compares With Proprietary Models

In the overall Artificial Analysis ranking, which scores open and closed models side by side, MiMo-V2.6-Pro comes in sixth. Every model ahead of it is proprietary:

• Claude Fable 5.1 (Anthropic): 53 points
• GPT-6 Astra (OpenAI): 53 points
• Claude Opus 5 (Anthropic): 51 points
• Muse Spark 1.3 (Meta): 48 points
• GPT-5.6 Sol (OpenAI): 47 points
• MiMo-V2.6-Pro (Xiaomi): 46 points

xAI’s Grok 4.7 is tied at 46 points, while Google’s Gemini 3.8 Flash (41 points) and DeepSeek V4.1 Pro (36 points) fall behind, as VentureBeat reports. That leaves a gap of seven points to the top of the field and a single point to the next proprietary model up.

The cost picture is where the model stands out most. Artificial Analysis puts the spend at roughly 0.11 euros per Intelligence Index task, placing MiMo-V2.6-Pro on the Pareto frontier of intelligence against price. Throughput is measured at about 125 output tokens per second, good for 12th place among 114 tested models. The benchmarking service flags verbosity as a weak spot: the model burned 140 million output tokens to complete the index run.

Architecture, Pricing, and Availability

MiMo-V2.6-Pro is a sparse mixture-of-experts model with 1.02 trillion total parameters, of which only 42 billion activate per inference. Flash carries 310 billion parameters with 15 billion active. Both handle text, image, speech, and video natively as input, produce text as output, and offer a context window of one million tokens.

API pricing stays level with the previous generation. Pro costs 0.435 US dollars (about 0.38 euros) per million input tokens and 0.87 US dollars (about 0.76 euros) per million output tokens, with a 99 percent discount on cache hits. Flash runs at 0.14 US dollars (about 0.12 euros) for input and 0.28 US dollars (about 0.24 euros) for output. The UltraSpeed variant costs ten times as much as standard Pro.

The models are available through Xiaomi’s AI Studio, MiMo Code, the MiMo Desktop app that just left early access, the company’s own API platform, and OpenRouter.

Strengths: Agents, Cybersecurity, and Visual Coding

The benchmark results Xiaomi published itself paint a nuanced picture. MiMo-V2.6-Pro sits at or near the top in these areas:

• Cybersecurity: Pro reaches 94.0 on CyberGym, while Flash takes first place with 95.1, ahead of DeepSeek V4.1 Flash (88.1) and GLM 5.3 (84.5).
• Workflow automation: On Automation Bench, Pro lands at 53.1, just behind DeepSeek V4.1 Flash (54.8) and ahead of GPT-6 Astra (52.0) and Claude Opus 5 (50.3).
• Economically relevant tasks: Pro scores 1,673 on Artificial Analysis’ GDPVal, third behind Claude Fable 5.1 (1,735) and Claude Opus 5 (1,708).
• Tool use: On Toolathlon-verified, Pro reaches 76.9, ahead of GPT-5.6 Sol (74.9) and narrowly behind the Claude models.
• Visual coding: Pro takes 72.3 on Xiaomi’s in-house visual coding test, above Claude Opus 5 (70.0) and below GPT-6 Astra (82.2).
• Software engineering: Pro hits 71.9 on DeepSWE v1.1, with DeepSeek V4.1 Flash (74.2), Claude Opus 5, and GPT-6 Astra (74.0 each) leading.

The model trails clearly where long terminal sessions are involved. On Terminal Bench 4.0 it reaches 34.9, against 59.6 for GPT-6 Astra and 49.0 for Claude Opus 5. ExploitGym tells a similar story, with Pro at 17.8, well behind the leading proprietary systems.

Reinforcement Learning at Scale, Streamed in Public

Xiaomi credits the leap to scaled reinforcement learning on verifiable, complex tasks. In under six days, Flash and Pro each completed 30 RL steps across roughly 750,000 trajectories, at a cost of around 740,000 euros for Flash and 2.28 million euros for Pro. On the held-out DeepSWE v1.1 benchmark, Pro climbed from 58.4 to 72.6 points and Flash from 48.8 to 65.7.

Training used 1,568 samples per update, a fully asynchronous architecture, and context lengths of up to one million tokens. The company says it streamed the production run live as it happened. Alongside the weights, Xiaomi is releasing the technical report, the training environments, and the RL code, so that researchers can reproduce and verify the results.

From Vibe Coding to Vibe World

Xiaomi’s headline pitch is that coding ability now generalizes to work well outside software engineering, something the company calls “Vibe World.” That covers generating 3D scenes with interaction logic from image, video, or text prompts, modeling 3D objects in Blender, and controlling a Franka Panda robotic arm in a simulation environment through multiple camera feeds.

Design and media applications follow the same pattern: frontend interfaces and slide decks from a one-line brief, video production including editing and scoring, plus first attempts at composition. In one demonstration, MiMo-V2.6-Pro wrote a score for an orchestra of roughly ten instruments and converted it to MIDI on its own.

Two case studies come from research. Working with Xiaomi’s materials experts, the model designed a new metal-organic framework for binding PFAS, reviewing literature and patents and computing binding strengths with open-source simulation tools. In a second case, it helped formalize the main theorem from Li and Yorke’s paper “Period Three Implies Chaos” in Lean 4, a project spanning more than 6,000 lines of code that the Lean kernel verified in full. According to the company, the model had received no Lean-specific post-training.

What It Means

MiMo-V2.6-Pro puts another Chinese open-weight model within striking distance of the global frontier, following pressure from Alibaba’s Qwen3.8-Max and Kimi K3. For developers and companies that want to self-host or fine-tune, the ceiling has moved on what is achievable without license fees and without tying themselves to a US provider. Trending Topics has looked at the role open models can play in Europe elsewhere.