Arcee's 400B Thinking Model and Liquid's Edge Titan Square Off in Open Source

Tools

Arcee Releases Trinity-Large-Thinking: Apache 2.0, 400B Parameters, Built in the US

This is the most credible open-weight challenger to proprietary reasoning models right now, scoring #2 on PinchBench behind Claude Opus-4.6 at roughly 96% lower cost per output token. The Apache 2.0 license and downloadable weights mean enterprises can actually own and host this, which is the gap Meta left wide open after Llama 4 stumbled.

Liquid AI's LFM2.5-350M Hits 40K Tokens Per Second on H100 and Runs on a Raspberry Pi

Trained on 28 trillion tokens with a hybrid LIV architecture, this 350M model beats rivals more than twice its size on instruction following and tool use benchmarks. If you are building edge pipelines or high-volume data extraction, the memory footprint numbers here are hard to ignore: 81MB on a Snapdragon GPU.

Analysis

Trinity-Large-Thinking Architecture Deep Dive: Sparse MoE, Muon Optimizer, and SMEBU Load Balancing

The technical details here are worth understanding if you are building on top of this model. Activating only 13B parameters per token via 4-of-256 expert routing means you get 400B-scale knowledge without dense inference costs, and the 262k context window makes it practical for long agentic loops.

LFM2.5-350M Technical Breakdown: Hybrid LIV Backbone and 80,000:1 Token-to-Parameter Ratio

The departure from pure Transformer architecture is the real story here. Using Linear Input-Varying Systems alongside a small number of GQA blocks cuts KV cache overhead dramatically, which is why this model can run on constrained devices where standard transformers would fall over.