Arcee Trinity-Large-Thinking vs Ternary Bonsai

Arcee AI released Trinity-Large-Thinking on April 2, 2026 — a 398 billion parameter sparse Mixture-of-Experts reasoning model under the Apache 2.0 license. Built by a 35-person startup that committed $20 million (nearly half its total funding) to a 33-day training run on 2,048 NVIDIA B300 Blackwell GPUs, it's one of the most ambitious open-source bets from a US AI lab. The architecture is unusually sparse: 256 experts with only 4 active per token (a 1.56% routing fraction), which delivers 2–3× faster inference throughput compared to dense models of similar parameter count. At $0.90 per million output tokens via the Arcee API, it costs approximately 96% less than Claude Opus 4.6 at $25 per million — while scoring within two benchmark points on key agent tasks. For enterprises that need a powerful model they can download, fine-tune, and deploy on their own infrastructure without licensing restrictions, Trinity-Large-Thinking fills a real gap. Apache 2.0 means no restrictions on commercial use, and the US origin is an increasingly relevant compliance factor for government and defense customers.

PrismML's Ternary Bonsai is a family of ultra-compressed language models using 1.58-bit weights — meaning every parameter is stored as -1, 0, or +1, with no higher-precision layers anywhere in the architecture. The line-up covers 8B, 4B, and 1.7B parameter models. The flagship 8B model fits in 1.75 GB of RAM, a 9x reduction versus a 16-bit baseline. Unlike earlier 1-bit experiments that felt like a party trick with serious capability regressions, Ternary Bonsai 8B outperforms PrismML's own prior 1-bit Bonsai 8B by 5 points on average across standard benchmarks. The team also ships WebGPU inference, so the 1.7B model runs entirely in a browser tab. This is the first time a production-quality chat model has run with no server at all. The real-world use case is edge and offline deployment: medical devices, air-gapped government systems, consumer apps that need to work without a signal. At 1.75 GB, the 8B model fits on the GPU RAM of a six-year-old gaming laptop. PrismML is positioning this as the foundation for truly offline AI — a credible claim if the capability benchmarks hold up under real-world testing.