Hugging Face: Same Model Scores 62% and 33% Under Different Harnesses
Hugging Face released a multi-harness RL guide, noting that the same model with identical weights can drop from a score of 62% to 33% under different agent harnesses. Black Forest Labs released Flux 3 Image, which supports multi-step editing without altering other parts of the image. Meta released six papers from collaborations between mathematicians and Muse Spark.
Model Releases/Updates
Black Forest Labs Releases Flux 3 Image, Supporting Multi-Step Editing
Black Forest Labs released Flux 3 Image, the image model in the Flux 3 model family. The company says it supports multi-step editing without altering other parts of the image, covering text-to-image, image-to-image, text rendering, and photorealistic generation.
Ai2 Open-Sources AstaBrief 8B Scientific Report Generation Model
Ai2 open-sourced AstaBrief 8B, a model that turns research questions and retrieved literature snippets into reports with citations. It is based on Qwen3-8B and trained with SFT and DPO, and is already live as Asta's Fast mode.
Product Releases/Updates
NVIDIA DGX Spark Launches 64GB Version, Supports Two-Unit Cluster Scaling for Local AI
NVIDIA will launch a 64GB unified memory version of DGX Spark this month through Acer, ASUS, Dell, Gigabyte, HP, and MSI. It goes on sale October 23, starting at $4,999, retaining the GB10 Grace Blackwell superchip, DGX OS, and the full NVIDIA AI software stack, and supporting models with up to 100 billion parameters running locally on the device.
Paper Research
Meta publishes six papers from collaboration between mathematicians and Muse Spark
Meta said that after its AI models achieved gold-medal-level performance in five competitions spanning mathematics, physics, and chemistry, it had mathematicians collaborate with the Thinking Mode of Muse Spark 1.1 and 1.2 through the meta.ai chat interface to solve open problems for which no ready-made path yet exists, without a custom research scaffold.
Google releases new-generation federated learning system based on TEEs, already adopted by Gboard
Google Research released a new-generation federated learning system that uses trusted execution environments (TEEs) to provide verifiable, auditable guarantees for data anonymization. Gboard has already adopted the system to launch next-word prediction models in English and Japanese, with stronger privacy guarantees and improved accuracy.
