The Next Top Open Model, Google Voice Agents, DeepSeek Shrinks Caches: Plus a letter on open weight cybersecurity capabilities
Summary
- Anthropic analyzed the cyber capabilities of open-weight model GLM-5.3, finding its exploit success rate approaches closed-weight Mythos.
- Xiaomi released MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash, top-scoring open-weight models on Artificial Analysis’ Intelligence Index and Vals Index.
- Xiaomi published over 7,000 reinforcement learning task environments and training code alongside its new multi-modal Mixture-of-Experts models.
- Xiaomi's training methodology used specialized AI quality graders to reward code that human reviewers would accept, not just code that passes automated tests.
Why it matters
The release of highly capable open-weight models and training recipes like Xiaomi's MiMo series lowers costs and advances agentic AI capabilities, while highlighting ongoing cybersecurity and code quality engineering challenges.