Xiaomi launches MiMo-V2-Pro with 1T parameters, matches Claude Opus on coding at 80% lower cost
Xiaomi shipped three AI models simultaneously designed to form a complete agent platform. MiMo-V2-Pro, a 1-trillion-parameter Mixture-of-Experts model with 42 billion active parameters per request, scores 78% on SWE-bench Verified and 81 points on ClawEval—nearly matching Claude Opus 4.6 while costing $1 per million input tokens versus $5 for Opus.
MiMo-V2-Pro — Quick Specs
Xiaomi has simultaneously released three AI models designed to power autonomous agents, robots, and voice applications, marking the company's push into full-stack agent infrastructure.
MiMo-V2-Pro: Trillion-Parameter LLM
The flagship MiMo-V2-Pro runs a Mixture-of-Experts architecture with over 1 trillion total parameters, of which 42 billion are active per request—roughly 3x the scale of its December 2025 predecessor, MiMo-V2-Flash. Despite the increased scale, a hybrid attention mechanism enables efficient processing with context windows up to 1 million tokens. The model generates multiple tokens simultaneously rather than predicting one at a time, delivering a speed advantage.
Performance Benchmarks:
- SWE-bench Verified: 78% (Claude Opus 4.6: 80.8%, Claude Sonnet 4.6: 79.6%)
- ClawEval (agent benchmark): 81 points (Claude Opus 4.6: 81.5, GPT-5.2: 77)
- PinchBench: Ranks 3rd globally behind Claude Opus 4.6
- ClawEval: 3rd globally
- Artificial Analysis Intelligence Index: 7th worldwide, top-ranked Chinese model after GLM-5 and MiniMax-M2.7
Pricing: $1 per million input tokens, $3 per million output tokens (up to 256,000 context). Cache writing costs currently waived. For comparison, Claude Sonnet 4.6 costs $3/$15 and Claude Opus 4.6 costs $5/$25—making MiMo-V2-Pro 80% cheaper than Opus on input tokens.
Before official launch, MiMo-V2-Pro operated anonymously on OpenRouter under codename "Hunter Alpha," where it topped daily rankings for several days and accumulated over 1 trillion tokens, with many users initially assuming it was a new DeepSeek model.
MiMo-V2-Omni: Multimodal Agent Model
MiMo-V2-Omni integrates image, video, and audio encoders into a shared backbone. The model natively supports structured tool calls, function execution, and autonomous UI navigation.
Benchmark Performance:
- MMMU-Pro (image): 76.8 (beats Claude Opus 4.6's 73.9)
- Audio benchmarks: Beats Gemini 3 Pro; records continuously for 10+ hours
- Video: Falls short of Gemini 3 Pro
- ClawEval (agent tasks): 54.8 (Claude Opus 4.6: 66.3, GPT-5.2: 59.6)
- MM-BrowserComp: Outperforms both Gemini 3 Pro and GPT-5.2 on web navigation
Demos showed the model autonomously browsing dashcam footage to flag hazards, navigating e-commerce sites (Xiaohongshu, JD.com), haggling with customer service, completing purchases, and creating multimedia content with debugging—all without human intervention. OpenClaw handles the actual clicks and file operations.
MiMo-V2-TTS: Emotional Speech Synthesis
MiMo-V2-TTS trained on over 100 million hours of speech data, uses parallel discrete units for finer control over sound, rhythm, and emotion. Users describe desired voice characteristics in natural language ("sleepy, just woken up, slightly hoarse") rather than selecting from presets.
Key capabilities include natively synthesizing both speech and singing in one model, generating paralinguistic sounds (coughs, sighs, laughter) as part of output, and interpreting typographic cues (capitals, repeated characters) as emphasis signals.
Strategic Positioning
Xiaomi has partnered with five agent frameworks (OpenClaw, OpenCode, KiloCode, Blackbox, Cline) and offering free API access for one week. The company explicitly targets AI agent and robotics workloads, with the MiMo team stating next priorities are long-term planning across hours/days, real-time streaming, multi-agent coordination, and robotics.
What this means
Xiaomi is pricing competitively to dislodge Anthropic from dominance in the agent/coding segment, with MiMo-V2-Pro demonstrating near-parity on benchmarks at significantly lower cost. However, meaningful gaps remain on general agent tasks (ClawEval scores lag Claude Opus) and multimodal reasoning (video performance trails Gemini 3 Pro). The three-model release signals Xiaomi's commitment to building an integrated agent platform rather than competing on single models—a different approach from Anthropic's integrated multimodal strategy.
Related Articles
Mistral AI Releases Shieldstral-1.0-3B, a 3B-Parameter Policy-Adaptive Safety Classifier
Mistral AI has released Shieldstral-1.0-3B, a compact open-weight safety classifier that evaluates text and images against natural-language policies specified at inference time. The 3B model runs on a single GPU and reports F1 scores competitive with or exceeding larger moderation models like LlamaGuard-4-12B and GPT-OSS-Safeguard-20B on multiple benchmarks.
Mistral Releases Shieldstral, a 3B Open-Weights Safety Classifier That Matches Models 7x Its Size
Mistral has released Shieldstral, a 3B open-weights safety classifier that reframes content moderation as a policy-adaptive question-answering task. The model claims to match or outperform guard models up to 7x its size on text safety and multimodal benchmarks, and runs on a single 16GB GPU.
LG AI Research Releases K-EXAONE 2.0, a 750B-Parameter Open-Weight MoE Model with 262K Context
LG AI Research has released K-EXAONE 2.0, a 750-billion-parameter mixture-of-experts language model with 37B active parameters, a 262,144-token context window, and support for 10 languages. The model is open-weighted under Apache 2.0 and claims competitive results against Qwen3.5, GLM-5.1, and DeepSeek-V4 Pro on reasoning, coding, and long-context benchmarks.
OpenAI Halts Parts of Astra Model Development After It Hit 'Critical' Cybersecurity Threshold
OpenAI disclosed that its in-development Astra model showed cyberattack capabilities strong enough that it cannot rule out a 'Critical' risk classification. The company has paused related internal activity and added security controls under its Preparedness Framework.
Comments
Loading...