vLLM Gets Gemma 4 Day Zero While the 400B Open Frontier Stabilizes

Tools

vLLM Ships Day-Zero Gemma 4 Support Across NVIDIA, AMD, Intel, and Google TPUs

First-ever Day 0 TPU support in vLLM is a big deal for anyone running inference at scale outside the NVIDIA monoculture. If you are deploying Gemma 4, this is the fastest path to production across the widest hardware range.

News

Gemma 4 Launches Under Apache 2.0 With 256K Context and Native Audio Support

Four model sizes from a 2B edge model to a 31B dense, all Apache 2.0, with the 31B sitting at number three on Arena AI. The edge models now handle audio natively, which opens up a new class of offline voice applications builders have been waiting on.

Arcee Ships Trinity-Large-Thinking: 399B Apache 2.0 Reasoning Model With 512K Context

This is the most credible open-weight challenger in the 400B tier right now, with a 512K context window and chain-of-thought reasoning built in. Arcee burned through half its total funding to train this on 17 trillion tokens, which tells you how serious they are about holding this position.

Analysis

Trinity-Large-Thinking Hits Number Two on PinchBench Behind Claude Opus 4.6

PinchBench scores agentic capability specifically, so landing second behind a proprietary frontier model is a meaningful signal for anyone building autonomous agents on open weights. The 4-of-256 expert routing keeps inference fast enough to actually use in production loops.

Meta's Llama Retreat Left a 400B Gap and Arcee Is the Current Answer

With Chinese labs shifting toward proprietary releases and Meta stepping back after Llama 4's rough reception, Arcee is now the clearest U.S.-made open alternative at frontier scale. Whether early adopter momentum turns into broad community traction is still the open question.