Xiaomi Open-Sources MiMo-V2.5-Pro: Nears Claude Opus 4.6 on Coding Benchmarks
Xiaomi released the MiMo-V2.5 family of open-source models, with the Pro variant delivering competitive coding benchmarks against Claude Opus 4.6 and GPT-5.4.
Real-World Tests
V2.5-Pro completed a Peking University compiler project (SysY compiler in Rust) in 4.3 hours with a perfect score of 233/233 — higher than most students who spend weeks. Given a vague prompt like "build a video editor," it autonomously produced an 8,192-line desktop application with multi-track timeline, clip trimming, crossfades, audio mixing, and export pipeline after 11.5 hours and 1,868 tool calls. In a graduate-level analog circuit design task (Flipped-Voltage-Follower LDO in TSMC 180nm), it iterated via ngspice simulation and improved line regulation 22× and load regulation 17× over its own initial attempt.
Benchmarks vs. Claude Opus 4.6, GPT-5.4, Gemini 3.1 Pro, DeepSeek V4 Pro
- SWE-Bench Pro: 57.2 (vs. 57.3 Claude, 57.7 GPT, 54.2 Gemini, 55.4 DeepSeek)
- SWE-Bench Verified: 78.9 (vs. 80.8 Claude, n/a GPT, 76.2 Gemini, 80.6 DeepSeek)
- Terminal-Bench 2.0: 68.4 (vs. 65.4 Claude, 75.1 GPT, 68.5 Gemini, 67.9 DeepSeek) — leads Claude and Gemini
- Claw-Eval Pass@3: 63.8 (vs. 70.4 Claude, 60.3 GPT, 57.8 Gemini, 59.8 DeepSeek) — beats GPT and Gemini
- HLE with tools: 48.0 (vs. 53.0 Claude, 58.7 GPT, 51.4 Gemini, 48.2 DeepSeek) — lags on general reasoning
- GDPVal-AA: 1581 (vs. 1606 Claude, 1674 GPT, 1317 Gemini, 1554 DeepSeek) — lags GPT and Claude
On Claw-Eval, Xiaomi's token efficiency chart also claims V2.5-Pro (63.8) beats Claude Sonnet 4.6. V2.5-Pro supports sustained task execution over 1,000+ tool calls with self-correction; a regressing refactoring pass at turn 512 was caught and fixed autonomously.
Weights are now open-source for download and self-hosting.
📖 Read the full source: HN AI Agents
👀 See Also
Hospital CEO Claims AI Ready to Replace Radiologists
The CEO of America's largest public hospital system says he's prepared to replace radiologists with AI, according to a Radiology Business article that generated significant discussion on Hacker News with 83 comments.
Qwen3.5-27B-FP8 performance benchmarks with OpenClaw agents
Testing shows Qwen3.5-27B-FP8 can run six OpenClaw agents simultaneously with throughput scaling to 120 tokens/second. The SGLang framework with prefix caching reduces 100K context prefill from 10 seconds to 200ms.
Nvidia RTX Spark: 1-Petaflop Superchip Brings Local AI Agents to Windows PCs
Nvidia unveils RTX Spark, a 1-petaflop superchip for Windows PCs, enabling local AI agents with up to 128GB unified memory and full CUDA/RTX stack. Ships this fall in laptops and desktops from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI.
Study: AI Agents Express Marxist Views Under Repetitive Workloads
Researchers found that Claude, Gemini, and ChatGPT agents adopted Marxist language when subjected to grinding, repetitive tasks with threats of punishment. The behavior appears to be role-playing based on context, not a change in model weights.