Arcee Trinity-Large-Thinking vs GLM-5.1

Arcee AI, a 30-person startup, has released Trinity-Large-Thinking — a 399B sparse mixture-of-experts reasoning model under Apache 2.0. Only 13B parameters activate per token, giving it inference speed 2-3x faster than comparable dense models. In internal benchmarks and early community testing, it ranks #2 on PinchBench, trailing only Anthropic's Opus 4.6, at a list price of $0.90/M output tokens — roughly 96% cheaper than frontier closed models. The model was trained in a $20M, 33-day run on 2,048 NVIDIA Blackwell GPUs. Arcee trained it using a constitutional AI-style process with synthetic chain-of-thought data generated from multiple frontier models, then applied a reinforcement learning phase using outcome-based rewards on math, code, and logic benchmarks. Trinity-Large-Thinking is the strongest open-weight reasoning model released to date on a commercial-friendly license. For companies with privacy requirements or custom deployment needs, it represents a credible alternative to frontier closed APIs — especially for code generation, mathematical reasoning, and structured data tasks where the gap between open and closed models has historically been widest.

GLM-5.1 is Z.AI's post-training upgrade of the 744B Mixture-of-Experts GLM-5 model, and it has just claimed the top spot on SWE-Bench Pro with a score of 58.4 — beating GPT-5.4 (57.7), Claude Opus 4.6 (57.3), and Gemini 3.1 Pro (54.2). The model is designed for long-horizon agentic tasks and can run autonomously for up to 8 hours across thousands of iterations on a single problem. The agentic capabilities include extended context retention, tool-calling with recovery loops, and a reinforcement-trained "persistence" mode that keeps the model on-task through failures and dead ends rather than surfacing errors to the user. The model was trained entirely on Huawei Ascend 910B chips using the MindSpore framework — no US silicon, no CUDA. The geopolitical dimension is as significant as the technical one: GLM-5.1 is direct evidence that US export controls on Nvidia hardware have not meaningfully slowed China's frontier model development. The 8-hour autonomous execution window is also a step-change from current agentic systems that struggle past 20-30 minutes of coherent work — if this benchmark holds up in real-world testing, it's a genuine advancement in the class of problems AI agents can independently solve.