Gemma 4 Goes Apache and Arcee Fills the 400B Gap Meta Left Behind

News

Google Releases Gemma 4 Under Apache 2.0 With Four Model Sizes

Gemma 4 ships in E2B, E4B, 26B MoE, and 31B Dense variants, all under Apache 2.0 with 140+ language support and up to 256K context. The 31B model is already ranked #3 on Arena AI's open model leaderboard, competing with models 20x its size.

Arcee AI Ships Trinity-Large-Thinking: 399B Apache 2.0 Reasoning Model

A 26-person startup spent $20M and 33 days on 2,048 Blackwell GPUs to ship a 399B sparse MoE model that activates only 13B parameters per token. It ranks #2 on PinchBench for agentic tasks, sitting just behind Claude Opus 4.6, and the full weights are on Hugging Face under Apache 2.0.

Analysis

Trinity-Large-Thinking Technical Breakdown: SMEBU, Muon Optimizer, and 262K Context

The architectural details here are worth reading if you are building agents: SMEBU load balancing prevents expert collapse in the MoE routing, and the Muon optimizer improved capital efficiency over AdamW during the 17T token pre-training run. The 262K context window and multi-turn tool calling focus make this a practical pick for long-horizon agentic workflows.

Why the 400B Open Model Gap Matters Now That Meta Has Stepped Back

Meta's Llama 4 launch in April 2025 was messy enough that the Llama division effectively retreated from the frontier, and Chinese labs are quietly moving toward proprietary releases too. That leaves Arcee as one of the few credible sources of a downloadable, customizable 400B-class model right now.

Tools

Gemma 4 Lands on vLLM With Day-Zero Support Across NVIDIA, AMD, Intel, and Google TPUs

Day-zero multi-backend support means you can run Gemma 4 on whatever hardware you have without waiting for downstream integrations. Native function calling, structured JSON output, and vision plus audio across the model family makes this immediately useful for production agent stacks.