A scrappy team just built a 400B open source LLM from scratch that beats Meta’s Llama on coding and math—developers, your new favorite toy is here.
You’ve been waiting for an open source giant that actually delivers on benchmarks without the Chinese model drama.
Arcee AI, a tiny startup, shocked the world by releasing Trinity, a 400B-parameter open source base LLM built entirely from scratch. It’s currently in preview with more post-training underway, but early benchmarks already show it holding its own—or slightly beating—Meta’s Llama 4 Maverick on coding, math, common sense, knowledge, and reasoning tasks.[5] They followed this with smaller siblings in December: 26B Trinity Mini for reasoning in web apps and agents, and 6B Trinity Nano for ultra-tiny chatty models.
This matters because developers and academics now have a U.S.-built alternative to woo companies away from Chinese open models. Multimodal expansions like vision and speech-to-text are coming soon, making it versatile for real apps. No more settling for underperformers when scaling agents or fine-tuning for production.[5]
Compare to Llama: Arcee’s base model edges it out in key dev benchmarks without full multimodality yet (Llama 4 has text+images). It’s not just hype—it’s targeted at your workflows, from inference speed to token efficiency.[5]
Grab the preview from their site, benchmark it against your stack, and watch for the full release. Could this finally give open source the edge to dominate proprietary giants?
Source: TechCrunch