Neurodigest for the Week (#116)

Open Models Weights of Xiaomi MiMo 2.5 released - a million context tokens, multimodality (even audio), decent agent results. Mistral Medium 3.5 - the team updated the existing pretrain. It shows very well why no one trains large dense models anymore. Laguna XS.2 - a decent model from Poolside AI, on par with Qwen 3.5 35B-A3B at a similar size, but not Chinese. Miscellaneous GPT 5.5 appeared on the coding arena - results show very well why the arena is trash. OpenAI added pets to Codex - the team is fooling around, while Codex's revenue doubled in a week. Talkie - an esoteric LLM trained on data up to 1930 inclusive.