AI
OpenAI fired three safety researchers as loop-transformer model hides its reasoning
OpenAI dismissed three monitoring researchers on October 1, as its new Astra architecture keeps some computations in hidden states that safety tools cannot read.
This edition was produced with artificial intelligence. Text and voice are generated automatically.
Meta releases Muse Spark Open, an 816B open-weight model built to run on Macs
Meta launched Muse Spark Open, calling it the most capable open-weight model in the United States, with 816 billion parameters and a 1 million token context window. Evaluation results released Thursday claim superior performance to other popular open-weight models. The model is natively multimodal and built for complex coding and agentic work, designed to run locally on Apple computers, though that requires at least 1 terabyte of RAM and works best on a Mac Studio with the M5 Ultra chip. Meta paired it with Muse Code, its AI coding terminal, and deployed it to tasks including a full Mastodon client build and footage review in DaVinci Resolve. It is available through several API platforms.
Meta launches an open-weight AI model that runs on Macs →
Topics: Meta Muse
Meta pitches Muse Spark Open against Qwen, GLM and Kimi, but not its paid models
Muse Spark Open carries 816 billion parameters with 42 billion active, supports up to one million tokens of context, and handles text, image, video and audio. Meta says it leads on JobBench, MCP-Atlas, HLE-Diamond and ZeroBench in several categories ahead of Qwen 3.8, GLM 5.3 and Kimi K3, but still falls well short of its paid Muse Wonder Full models and trails some rivals on coding benchmarks. It is optimized for Apple’s M5 Ultra and runs locally on two Mac Studios with 512 GB of unified memory each. Weights are on Hugging Face under a free license requiring the label Built with Muse Spark Open; providers earning over $20 million annually and offering it as an API need a separate agreement.
Muse Spark Open is Meta’s answer to Qwen, GLM, and Kimi in the open-weight race →
Topics: Meta MuseHugging Face
Black Forest Labs ships FLUX 3 Image Klein for GPUs with 8 GB of VRAM
Black Forest Labs released FLUX 3 Image Klein, a 5.6-billion-parameter open-weights model for text-to-image generation, editing and generation with up to ten reference images. It supports native 4K output, pixel-accurate editing and bounding boxes for element placement. The company says it scores the highest Elo ratings among open models in internal testing, beating Ideogram V4.5 and its own FLUX.2 [pro], and gains 91 Elo points over FLUX.2 [klein] with 38 percent fewer parameters. An INT8 build runs on NVIDIA GPUs with just 8 GB of VRAM, generating an image in under 9 seconds on an RTX 4070. Weights ship under the FLUX Kommunity License, free for commercial use by companies under $5 million in annual revenue, and the model is also available via the BFL API at $0.009 per 1,000 images.
Nvidia launches Nemotron 3.5 Super VL for always-on enterprise agents
Nvidia released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters built for enterprise agents running continuously in the background on text, images, documents and videos. It uses a Mixture-of-Experts architecture with 12 billion parameters active per token and was developed with the Nemotron Coalition, including Mistral AI, Thinking Machines Lab, Perplexity and Black Forest Labs. Two compact Bolt models launch alongside it, including Nemotron 3.5 Bolt 74B-A9B 24GB for local agent tasks. Nvidia says Super VL trails leading closed models by a wide margin in the Artificial Analysis ranking but beats its own Nemotron 3 Ultra, and on a single GPU delivers up to five times the throughput of similar-sized open models on agent tasks. Weights, datasets and training recipes are on Hugging Face.
Nvidia’s Nemotron 3.5 Super VL is an open model built to run enterprise AI agents nonstop →
Topics: NvidiaHugging Face
OpenAI fired three researchers as Astra hid reasoning from safety monitors
OpenAI dismissed three researchers on October 1: Tomek Korbak, Mikita Balesni and Jasmine Wang. All had worked on monitoring OpenAI’s models and had publicly criticized the company’s direction. OpenAI said they violated policies on sensitive information and denied the firings concerned safety or dissent. Balesni wrote on X that he believes they were fired for prioritizing safety over near-term corporate interests. The Information reported on October 13 that the model meant to succeed GPT-6 Astra may extend the loop-transformer architecture, in which computation stays in hidden states never decoded into text, so chain-of-thought monitors cannot read it. Astra’s system card disclosed decreased chain-of-thought monitorability, and OpenAI called the issue manageable.
OpenAI Fired Three Researchers Who Said Loop Transformer Would Blind Safety Monitors →
Topics: OpenAIGPT-6 Astra