Open Source Finds New Anchors as Meta Retreats and Arcee Steps Up

News

Arcee Trinity-Large-Thinking Fills the 400B Open Model Gap Left by Llama 4

With Meta stepping back after Llama 4's rough reception, Arcee shipped a 400B open-weight reasoning model under a permissive license built entirely in the US. On agentic benchmarks like PinchBench (91.9) it's within spitting distance of Claude Opus 4.6 (93.3), which makes this a serious option for teams that need a large downloadable model they can actually customize.

Meta Launches Muse Spark as a Closed Proprietary Model, Ditching Open Weights

Meta's first model from its new superintelligence team is natively multimodal and leads on medical reasoning benchmarks, but the weights stay locked inside Meta. For builders who relied on Llama, this is the clearest signal yet that Meta is no longer a reliable open-source partner.

Tools

Google Releases Gemma 4 Under Apache 2.0 with Four Model Sizes and 256K Context

Gemma 4 ships in four sizes from a mobile-first E2B up to a 31B dense model, all under Apache 2.0 with no usage restrictions. The 31B is currently ranked third on the Arena AI open model leaderboard and runs on a single H100, making it a practical fine-tuning target for production teams right now.

Liquid AI Pushes LFM2.5-350M to 28T Tokens and Adds Large-Scale RL

LFM2.5-350M went from 10T to 28T training tokens and added reinforcement learning, and the results show up clearly in tool use (BFCLv3 22.95 to 44.11) and instruction following (IFBench 18.20 to 40.69). It hits 40K output tokens per second on a single H100 and runs on a Raspberry Pi 5, which makes it a real option for high-volume data pipelines or constrained edge deployments.