AMD published Instella-MoE-16B-A3B, a fully open Mixture-of-Experts language model trained from scratch on Instinct MI300X and MI325X GPUs (MarkTechPost). The model holds 16B total parameters but activates only 2.8B per token, using Gated MLA and FarSkip-Collective. AMD published weights from every training stage, plus data and training logs, making the release a full reproducibility artifact for open-weight MoE research.
NVIDIA AI released Molt, a PyTorch-native agentic reinforcement learning framework designed to reduce implementation friction in RL research (MarkTechPost). Molt composes Ray, vLLM, and NeMo AutoModel around a single asynchronous loop, replacing the typical pattern of threading algorithm changes through trainer, distributed backend, and rollout glue. The framework comprises approximately 8.6K lines of RL code. Separately, Supabase open-sourced supabase/evals, a benchmark and framework that runs coding agents including Claude Code, Codex, and OpenCode against real Supabase database tasks inside containerized stacks with deterministic scoring (MarkTechPost).