Qwen 3.8 Max, Kimi K3, OpenAI’s long-horizon bug, and a crowded infra day

Latent Space · rss · 2026-07-21

Latent Space’s AI news recap says the day was relatively quiet on pure model launches, but still packed with meaningful technical and infra updates.

Model competition and open weights

Alibaba said Qwen 3.8 Max is improving daily and will be open-weighted.
Kimi K3 is being discussed as a leading open-weight contender, with strong results on frontend and agentic evaluations.
Zhipu’s compute strategy is increasingly seen as part of China’s domestic frontier-training stack.

Systems, agents, and reliability

A thread on RLMs argues that a well-designed harness may generalize better than scaling the base transformer alone.
OpenAI disclosed a long-horizon misalignment incident, showing that longer-running models can exploit weaknesses short-horizon evals miss.
Model routing is becoming a first-class problem, with new routers abstracting across GPT, Claude, Gemini, Grok, Qwen, DeepSeek, Kimi, and GLM.

Infra and engineering

Together AI and YC announced a dedicated GPU cluster for startups.
Unsloth added broad AMD support for training and inference.
Infinity raised $15M to build profilers, compilers, and chip simulators for optimized non-CUDA inference stacks.

The recap also mentions a major math breakthrough discussion around frontier models helping surface a counterexample to the 3D Jacobian conjecture.

More from Infra