Neurodigest for Two Weeks (#118)

Neurodigest for Two Weeks (#118) LLM - Opus 4.8 — The model has become more honest, rarely cutting corners and better recognizing when it doesn't know something. The new low mode sometimes outperforms the old max, and the fast version has become three times cheaper. - MiniMax M3 — The M3 model with a million tokens of context has been released. It is currently available at a discount in the API and for free in OpenCode, with weights promised soon. - Gemma 4 12B — Google has released Gemma 4 12B, an open multimodality model without encoders. This hybrid reasoner has 256k context (Apache 2.0 license) and can process video, audio, and images through simple linear projections. - MAI-Thinking-1 — Microsoft has published a rare detailed technical report on the training of MAI-Thinking-1. It will not be open-sourced but will provide an API for fine-tuning. Generative Models - Wonders of Extreme Quantization — The startup PrismML has compressed FLUX.2 Klein 4B to 1 bit. The diffusion transformer now weighs only 930 MB and generates images directly in the browser or on iPhone. - Legal Neuro-Remixes and $9M Investments — My friends from the startup GRAI are building an AI music lab. They are currently actively hiring ML and Research engineers in Warsaw or remotely. Other - Open-source AI launcher from PewDiePie — PewDiePie has released Odysseus for self-hosting neural networks. The UX is on par with ChatGPT, but locally: with agent mode, Deep Research, and a built-in Cookbook. - New mega-client for SpaceX data centers — Google will rent 110,000 Blackwell servers from Musk for $920 million a month. Annually, the data centers will bring Musk about $26 billion.