Gemma 4 and Trinity's Deployment Stories Shift from Launch Hype to Real Builder Workflows

News

Arcee's Trinity-Large-Thinking: How a Scrappy Team Built a Frontier 400B Open Reasoning Model

Arcee shipped Trinity-Large-Thinking under Apache 2.0 with a full account of the tradeoffs, from 2048 B300s for pretraining to NVIDIA Dynamo for inference. At $0.90 per million output tokens and a #2 PinchBench ranking, this is the clearest signal yet that open weights can compete at the frontier without the frontier price.

Gemma 4's Four-Size Lineup Targets Everything From Jetson Orin to H100 Workstations

Google released Gemma 4 with four models sized for real hardware constraints, from 2B edge devices to a 31B dense workstation model, all under Apache 2.0 with 256K context. The 31B is ranked #3 among open models on Arena AI, and the 26B MoE activates only 3.8B parameters per inference pass, making it genuinely fast on consumer GPUs.

Tools

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

vLLM supporting Gemma 4 on day zero across four hardware backends is a bigger deal than it sounds: it means you can deploy immediately on whatever you already own without waiting for framework catch-up. First-ever day-zero TPU support in particular signals that the open inference stack is maturing fast.

Analysis

Running Always-On AI Agents Locally: What Zero Token Cost Actually Means in Real Numbers

This piece puts concrete math behind the local vs. cloud cost argument: 3 million tokens per month per background agent adds up fast at GPT-4-class rates. The hardware tradeoffs are laid out clearly, with Gemma 4 27B via Ollama as the practical starting point for most builders.