Meta releases Muse Spark Open, an 816B open-weight multimodal model
Meta published Muse Spark Open on 2026-10-15: a mixture-of-experts model with about 816B total and 42B active parameters, 256 experts per layer with 8 chosen per token, and a context window of up to about 1M tokens shared across text, image, audio and video input (output is text). It is aimed at agentic coding. Weights ship in BF16 (about 1.5 TB) and NVFP4 on Hugging Face, with community GGUF and MLX quantizations, and day-one support in Transformers, llama.cpp, vLLM and SGLang; it is also listed on OpenRouter. Meta's launch post calls it the most capable US open-weights model to date but gives no benchmark scores, and neither it nor the release notes state a license, so the claim is untested.
Sources & depth
NVIDIA releases Nemotron 3.5 Super VL, a 120B open multimodal model
NVIDIA's 120B-parameter mixture-of-experts vision-language model activates 12B parameters per token and targets agentic workflows. NVIDIA reports 80.8% on CharXiv chart reasoning (84% with a Python interpreter), 63.44% on Hermes Bench and 87.05% on LangChain Evals, plus up to 5x the throughput of open models of similar size on one GPU; these are NVIDIA's own figures. Base, post-trained and NVFP4 checkpoints, training recipes, data and RL environments are published, and the blog post does not name a license. Ollama v0.40.5 ships it as nemotron-3.5-super with up to 1M tokens of context, alongside a smaller 74B (9B active) Nemotron 3.5 Bolt sized for a 24 GB GPU.
Sources & depth
Simons Foundation, Anthropic and OpenAI launch a $15M effort for mathematicians
The Simons Foundation announced $15 million in combined funding and resources from itself, Anthropic and OpenAI to help mathematicians adapt to AI. The foundation will run an open call for proposals and support workshops on assessing AI-assisted results, learning the tools and developing shared standards for AI use; Anthropic and OpenAI will also provide tools and compute to researchers the foundation supports, including in science beyond mathematics.
Sources & depth
Black Forest Labs' FLUX.3 Image Klein arrives in Diffusers
FLUX.3 Image Klein, an open image model of roughly 6B parameters (the sources disagree between 5.6B and 6B), is supported in Diffusers from day one. A step-distilled variant makes a 1024x1024 image in about 2 seconds on an H100 with 16 steps and no guidance, and a base variant is provided for fine-tuning. It supports text-to-image, multi-reference editing and bounding-box layout control, with bf16, fp8 and NVFP4 weights and LoRA training scripts. The post does not state a license.
Sources & depth
Adobe and NVIDIA expand their partnership on GPU-accelerated creative tools
Adobe says Photoshop, Premiere and Substance 3D are optimized for NVIDIA's RTX Spark superchip, which launches 2026-10-16, with up to 2x faster AI processing, editing, color and effects. Adobe's Frame.io cloud pipeline reports about 5x faster embedding generation in early workloads, and Adobe Brand Concierge, now optimized with NVIDIA Nemotron models, reaches enterprise customers later this year. The figures are the companies' own.
Sources & depth
Transformers v5.20.0 adds six models and several breaking changes
Besides Muse Spark Open, the release adds Ling 3.0 Flash, Aleph Alpha's Kolibri 1, MiniCPM-V 4.7, Baidu's 3B UnlimitedOcr, UltraBERT encoders and MOSS-Transcribe-Diarize. Breaking changes include removal of deprecated cache utility methods, a minimum kernels version of 0.17.2 and a known speed regression for inference with expert parallelism, for which a fix is in progress.
Sources & depth