Anthropic claims its in-house Mythos model automated the porting of Claude to AMD GPUs in approximately three days, with minimal manual intervention (The Decoder). The effort is part of a broader strategy to run Claude across four chip families and reduce both infrastructure costs and vendor dependency. The company framed the achievement as evidence that AI tooling itself is eroding the historical difficulty of breaking into Nvidia’s CUDA ecosystem, a task that typically required months of engineering work.
Separately, MBZUAI’s Institute of Foundation Models released PAN 2, an open-weight 29-billion-parameter world model coupling an LLM planner with a diffusion transformer to simulate up to 20 seconds of action outcomes as video (MarkTechPost), available under the Apache 2.0 license. Fish Audio released Live 3, a full-duplex voice model for voice agents that listens while speaking, separates interruptions from vocal filler, and reacts in 186 milliseconds (MarkTechPost). Additionally, Meta FAIR shipped e-means in FAISS 1.16.0, a k-means replacement reaching convergence 10 to 45 times faster on GPU with no additional hyperparameter tuning (MarkTechPost).