Kyutai’s 2.4B REKI lets researchers and robot builders run a vision-language-action model trained only on public data, with performance close to roughly twice-sized policies on three benchmarks. It caches 32 small register tokens from recent camera frames inside the action expert, so each frame is encoded once instead of adding a separate memory module. That cut pose-oscillation failures by about fourfold, and Kyutai released the data, training, inference, evaluation, deployment code, and model weights.
Span-02 lets teams set a millisecond budget and get the model’s best decision so far, making variable-latency reasoning usable in production. Its budget can be fixed or chosen by an automatic gate, while the model outputs numeric answers with confidence rather than only yes or no. Respan lists hosted pricing at $0.01 per million input tokens with free output, and says it ran a browser-use race in 20.6 seconds versus GPT-6 Luna’s 38.5 seconds.
Tencent’s Hy-Memory-30B-A3B gives developers an open memory writer that scores above Claude Opus 5 across four evaluation harnesses, 55.49 to 54.62 on average. It builds a memory bank from an empty state, then carries the selected bank across context windows while reinforcement learning scores each chunk’s contribution. The comparison uses fixed prompts, retrieval settings, answer model, judge, and 20% benchmark subsamples, so the lead is narrow but directly measured.