Neurodigest for the Week (#99)

Neurodigest for the Week (#99) LLM - Gemini 3 Flash has been released — It outperforms 2.5 Pro, surpassing even Gemini 3 Pro in some benchmarks while being significantly cheaper. A hybrid reasoning model, it is already available in API and Google products. Generative Models - FLUX.2 [max] — It falls short of Nano Banana Pro but can work well with 10 references. - GPT Image 1.5 — Improved realism and added reasoning similar to Nano Banana Pro. A full-fledged alternative has been released. - SAM-Audio: a find for spies — Meta has added audio modality to SAM. Now you can isolate an object in a video and hear only its sound, separating it from the surrounding noise. - Kandinsky 5.0 Video in the text-to-video arena — The Pro version has become the top open model in the world in the T2V arena, falling short of SOTA models but on par with Luma Ray 3. Others - mini-SGLang — An optimized and minimalist inference engine (~5000 lines of code) for educational purposes and research, supporting the main features of SGLang. - Optimization of MoE training by DaoLab — The Dao AI Lab team is working magic again. SonicMoE is almost twice as fast and memory-efficient as the best open kernels for training MoE.