Arcee AI shipped Trinity Large Thinking on April 1, the reasoning-optimized checkpoint in their Trinity family. The release is worth more than a quick look for a few reasons, starting with the fact that a 30-person company trained a 400B-parameter model from scratch and is giving it away under Apache 2.0.
The architecture
Trinity Large uses sparse Mixture-of-Experts with 256 experts per layer, of which 4 are active at any given token. Total parameter count is ~398B; active compute per token is ~13B. The base model was trained on 17 trillion tokens across web, code, math, and reasoning data, running on 2,048 Nvidia B300 GPUs — which Arcee calls the largest publicly stated pretraining run on B300 machines.
The sparsity ratio here is extreme. Most MoE deployments run something like top-8 of 64 experts; Trinity’s 4-of-256 means roughly 1.5% of experts fire per token. The upside is inference throughput: Arcee claims 2–3x faster generation than comparable dense models. The downside is that routing at this level of granularity is tricky to get right — you need strong load balancing or a handful of popular experts eat all the traffic while most sit idle. The team used momentum-based expert load balancing and z-loss regularization to keep routing stable during training.
Context support is 512K tokens, with a pretraining window of 8K extended during mid-training.
Preview versus Thinking
The earlier Trinity Large Preview was a lightly post-trained instruct model — useful, but noticeably weaker on multi-step agent tasks. The Thinking variant adds extended chain-of-thought post-training with a stronger focus on multi-turn tool calling and long-horizon coherence. Arcee says it’s “far better than Preview at multi-turn tool use, context coherence, and instruction following across long-horizon agent runs.”
On PinchBench — Kilo’s benchmark for agentic model capability — Trinity Large Thinking ranks #2, just behind Claude Opus 4.6. The cost differential is roughly 96%: Thinking runs at $0.90/M output tokens versus Opus 4.6’s pricing. For agentic workloads where token counts pile up across long sessions, that gap matters in practice.
Why the license matters
Apache 2.0 is important here. The Trinity base model was trained by a team that includes DatologyAI and Prime Intellect as collaborators, and the weights are on HuggingFace in several quantizations (FP8, W4A16, GGUF). Permissive licensing means you can fine-tune it, run it behind a commercial API, or use it as a base for domain-specific post-training without navigating the custom license terms that trip up otherwise capable open models.
For comparison: the Trinity Large Preview had already served 3.4 trillion tokens on OpenRouter by the time Thinking dropped, making it the most-used open-weight model in the U.S. by that metric. The Thinking variant is likely to accelerate that, since agentic use cases are where token volumes grow fastest.
What to take seriously here
A 30-person startup training a 400B model from scratch is not a normal thing. The capital required — 2,048 B300 GPUs, reportedly costing around $20 million — is real. The technical report is substantial. And the benchmark position is legitimate: #2 on an agentic benchmark with a credible methodology, not a cherry-picked academic task.
What’s less clear is the Thinking variant’s full benchmark profile. Arcee hasn’t published a complete results table for it specifically; the numbers available are PinchBench and qualitative comparisons. MMLU, GPQA-Diamond, and math reasoning numbers for the Preview checkpoint are out there (GPQA-Diamond 63.3, AIME 2025 24.0), but the Thinking variant’s reasoning lift on formal math benchmarks hasn’t been published in detail yet.
Still: a fully open, commercially usable, 400B reasoning model at $0.90/M output tokens is a meaningful data point. The gap between what you can do with closed-API models and what you can do with open weights keeps narrowing.