Trinity-Large-Thinking: Scaling an Open Source Frontier Agent

TL;DR

Arcee AI releases Trinity-Large-Thinking, the reasoning post-trained variant of their 400B-total / 13B-activated sparse MoE Trinity Large base (see Trinity Large: An Open 400B Sparse MoE Model), under Apache 2.0 with API and Hugging Face weights. It is the official follow-up to the January Trinity-Large-Preview “instruct” model, with two months of additional SFT + RL pipeline work targeting multi-turn tool calling, context coherence, and long-horizon agent loops rather than raw benchmark coverage. Post-training was done on 1152 H100s; inference is served on Blackwell Ultra via NVIDIA Dynamo + vLLM. Headline claim: #2 on Kilo’s PinchBench (an agent-task benchmark used in their OpenClaw collection) behind Claude Opus-4.6 at $0.90 per million output tokens, ~96% cheaper than Opus.

Key claims

Trinity-Large-Thinking is the reasoning-tuned checkpoint on top of the Trinity Large 400B-total / 13B-activated MoE base announced in January, using a “thinking before responding” pattern previously seen in Trinity-Mini [§Trinity Large].
Post-training emphasized multi-turn tool calling, context coherence, instruction following, and stable behavior across long-running agent loops — explicitly positioned over benchmark-coverage breadth or “best coding model” framing [§Trinity Large].
Preview (the January light-instruct checkpoint, see Trinity Large: An Open 400B Sparse MoE Model) is reported as having served 3.37 trillion tokens on OpenRouter in its first two months, and listed as OpenRouter’s #1 most-used open model in the U.S. and #4 globally on OpenClaw [§Trinity Large].
Post-training infrastructure: 1152 H100 GPUs for the SFT+RL pipeline (separate from the 2048 B300 GPUs used for the 17T-token pretraining in January) [§Trinity Large].
Production inference stack: NVIDIA Dynamo + vLLM on Blackwell Ultra GPUs [§Trinity Large].
Scores #2 on PinchBench (Kilo’s agent benchmark, drawn from OpenClaw task workloads) behind Claude Opus-4.6 [§Trinity Large].
API pricing: $0.90 per million output tokens — Arcee frames this as roughly 96% cheaper than the leading frontier comparison [§Trinity Large].
Weights are released under Apache 2.0 on Hugging Face; the API is served via Arcee’s platform and OpenRouter. Trinity-Large-Preview remains free on OpenRouter on reduced hardware [§Trinity Large].
Authors flag a forward plan: distill Trinity-Large into successor Trinity-2-Nano and Trinity-2-Mini models — citing the “train a really good big one first, then distill” pattern [§A personal note].

Method

The post says relatively little about the post-training method itself beyond the high-level shape. Trinity-Large-Thinking is positioned as the same Trinity Large base architecture (400B-total / 13B-activated, 4-of-256 routing, six dense layers, Muon + SMEBU per the January base post — see Trinity Large: An Open 400B Sparse MoE Model) followed by an SFT + RL pipeline scaled over two months to match the 400B-parameter base. The model emits a “thinking” trace before its final response, the same pattern used by the smaller Trinity-Mini. The post emphasizes that this pipeline was scaled up from what worked at smaller scale on Trinity-Nano/Mini rather than re-architected.

Training and serving infrastructure: pretraining on 2048 B300s (carried over from January); post-training on 1152 H100s; production inference on Blackwell Ultra via NVIDIA Dynamo and vLLM.

Results

Headline numbers from the post:

PinchBench (Kilo’s agent benchmark, OpenClaw workloads): Trinity-Large-Thinking #2, behind Claude Opus-4.6 only.
API price: $0.90 / 1M output tokens — Arcee’s framing is “~96% cheaper than Opus.”
Preview-era traction (reported as motivation for the release): 3.37T tokens served via OpenRouter in the first two months of Preview; #1 most-used open model in the U.S. on OpenClaw, #4 globally.
Hardware: 2048× B300s (pretrain, prior phase), 1152× H100s (post-training), Blackwell Ultra (inference) with NVIDIA Dynamo + vLLM.

No tables, no per-benchmark comparison against open peers (GLM-4.5, DeepSeek V3.x, Kimi K2.5, etc.), and no released evals on tool-use benchmarks such as BFCL / GAIA / BrowseComp that the other reasoning-RL filed papers report. The agent-quality claim rests on PinchBench plus the OpenRouter usage data.

Why it’s interesting

This is the explicit reasoning-variant follow-up to the Trinity Large base model already filed at Luma (Trinity Large: An Open 400B Sparse MoE Model) — the January post explicitly flagged that “a reasoning-tuned variant is in post-training,” and this is that release. Together the two posts give a complete release-timeline datapoint for a 400B-class open MoE: 17T-token pretrain on 2048 B300s → light-instruct Preview ships with the base → two months of SFT+RL on 1152 H100s → reasoning checkpoint ships. For Open foundation-model releases this fills out the “shipping multiple checkpoints from one training run” pattern that Trinity established with Base / TrueBase / Preview, now with a fourth (Thinking) layered on top.

The agent-first framing puts it alongside MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling and Close the Loop: Synthesizing Infinite Tool-Use Data via Multi-Agent Role-Playing as another datapoint in Tool-Use Agents where the post-training target is multi-turn tool calling with long-context coherence rather than benchmark sweeps — but unlike those, Arcee discloses essentially nothing about the recipe (no GRPO/DPO mention, no reward shape, no rollout/trajectory-synthesis details). It contrasts with the Kimi K2.5 PARL recipe and MiroThinker’s three-stage SFT→DPO→GRPO pipeline in Reasoning RL precisely by being the dense / opaque counterpart — a useful negative datapoint for “what an open release looks like when it ships the weights but not the recipe.”