Mark McQuade

Today we drop Trinity-Large-Thinking. SOTA on Tau2-Airline, frontier-class on Tau2-Telecom, and the #2 model on PinchBench, right behind Opus. On BCFLv4, we're in the mix with the best. 26 people with under $50M raised and a ruthless pursuit of greatness. What this team just pulled off is nothing short of incredible. One hell of an accomplishment and I couldn't be more proud of Arcee. And we've got more to prove.

Arcee AI releases Trinity-Large-Thinking, a 399B-parameter MoE AI model under an Apache 2.0 license, allowing full customization and commercial use (@carlfranzen / VentureBeat) venturebeat.com/technology/arc… #a260403p1" target="_blank" rel="nofollow noopener">techmeme.com/260403/p1#a260… 📥 Send tips! techmeme.com/contact

Arcee x @NousResearch Trinity x Hermes Agent

Arcee's new, open source Trinity-Large-Thinking is the rare, powerful U.S.-made AI model that enterprises can download and customize venturebeat.com/ai/arcees-new-…

Display the api call for any prompt executed in chat. Copy the call directly into your terminal & update with your key to get started.

Jensen showcased PinchBench (by @kilocode) on stage at NVIDIA GTC as the new standard for evaluating @openclaw agent capability. Trinity-Large-Thinking just hit #2 on @pinchbench globally (91.9%) behind only Claude Opus 4.6 (93.3%), which costs ~20x more per token. That's an expensive percentage point... To celebrate the ongoing Trinity partnership, @OpenRouter is running a promotion to make Trinity-Large-Thinking free to use for OpenClaw until Sunday, April 5th. Even when the promo ends, the Trinity economics are pretty incredible. Trinity is $0.25/m input tokens (OpenClaw uses A LOT of input tokens), but with a 60%-90% cache hit rate at $0.06/m cache tokens the average input cost nets out to $0.087/m. You no longer have to compromise on logic to keep your agent infrastructure affordable. We're excited to see what you build.

Inference is evolving fast—and so are the models powering it. With @arcee_ai's Trinity, we’re seeing a new class of open-weight reasoning models: massive capability, radically more efficient economics. 👉 #2 on PinchBench (Kilo’s benchmark for agentic capability) 👉 ~96% lower cost than the top model That changes the equation. Now you can run advanced reasoning + agentic workloads on @digitalocean. at scale. This is the next wave of AI: thinking + doing, in one place. Check it out 👇 digitalocean.com/blog/run-advan… #AI #Inference #AgenticAI #OpenSource #DigitalOcean #LLMs

Now in Public Preview: @arcee_ai's Trinity Large-Thinking. ✨ The #1 U.S. model on OpenRouter. 3.4T tokens processed. 400B parameters. Live on DigitalOcean’s Serverless Inference. 🌊

Everyone is talking about Kimi and Qwen, but I'm honestly surprised more people aren't talking about models like Trinity from Arcee. I've been doing a deeper dive here and it's pretty interesting, here's a few differences that I'm not sure ppl fully realize. - Qwen and Kimi both have Apache licenses with restrictions in them. I'm honestly shocked most people don't know this. Once you hit over 100M monthly users, things change. - Arcee's models like Trinity use Apache 2.0 with NO restrictions, none, get over 100M monthly users, you're still in a good place. - The training data for Qwen and Kimi is undisclosed, for Arcee it's both disclosed and legally vetted. - One genuinely unique thing Arcee does: they released Trinity-Large-TrueBase, a raw 10-trillion-token checkpoint that hasn't undergone learning rate anneals or instruction tuning, letting researchers in regulated industries start from scratch for authentic audits and custom alignment. As I've been tinkering around with more small(ish) local LLMs, I've been doing a deeper dive, and honestly, I think I'm going to be doing more with Arcee vs. Kimi or Qwen because I can't help but think about scale. If I build something, I'd love to think that some day hundreds of millions, or billions of people, might use it, and I don't want to suddenly be dealing with license restrictions in this case and need to rearchitect. That being said, still pretty new to the Arcee models so I'll have to see how they perform. More to come, as always, I have a lot more to learn, and will share what I learn on here.

A big advantage of open weights is that companies can adapt models to their users. Cursor not disclosing its base models is frustrating, but it is not “embarrassing”, and it does not diminish how impressive its RL scaling has been in such a short time.

Thank you @arcee_ai

We’re looking forward to the @PrimeIntellect Day event tomorrow (3.14) in SF (always love a good Pi pun). Arcee CTO @latkins will be giving some updates on Trinity and the realities of scaling large-scale pre-training. He’ll be speaking alongside an impressive lineup of builders and researchers from Perplexity, Cognition, Hugging Face, and the team at Prime Intellect. The venue is officially booked out, but the waitlist is still open. If you’re in the Bay Area, it's worth jumping on the list in case a spot opens up. We’d love to see you there. luma.com/zjpepfzn

OpenClaw exploded into the ecosystem late January, right around the exact same time we launched Trinity Large Preview. It’s been incredible to watch these two projects scale in parallel over the last few months. Looking at the latest @openclaw usage on @OpenRouter it’s surreal to see Trinity Large Preview processing more volume than heavyweights like Claude, Gemini, GPT, and Grok. A huge thank you to the community for trusting Arcee as the model layer for this inventive new agentic crustacean.🦞

We were honored to be among the select few private beta partners for this release. Storage Buckets deliver fast, scalable, and interoperable object storage without the limitations of a git-backed system. We’ve loved kicking the tires on this over the last few weeks. Congratulations to the HF team on a massive launch!

Lineup @bratton (Antikythera / Berggruen Institute): The Singularity Will Not Be Singular @alexgraveley: Perplexity Computer @adityagrover_ (Inception): Diffusion LLMs @silasalberti (Cognition): Training Agentic Coding Models @Thom_Wolf (Hugging Face): Open Source Robotics @LucasAtkins7 (Arcee AI): Trinity & Scaling Large-Scale Pre-Training @KexinHuang5 (Phylo): Automating Science @vincentweisser: The Prime Intellect Masterplan @johannes_hage & @jannik_stra: The Prime Intellect Stack @samsja19: Scaling Long-Horizon Agentic RL @willccbb: Building Blocks for Synthetic Environments @a1zhang: Recursive Language Models

Trinity Large Preview just crossed 2T tokens in its first 35 days on @OpenRouter. We are grateful to our launch partners and core apps driving this traffic at @cline, @kilocode, @opencode, and @openclaw. Growth at this scale requires world-class infrastructure, and we couldn’t serve this volume without the continued support and reliability of our partners at @PrimeIntellect and @modal. Here's to the next 2T 🥂

Local adoption of Trinity Large is growing. While we expect high volume on our endpoints, it’s incredible to see a model of this size climbing to nearly 1,000 downloads per day on @huggingface. To those building with Trinity: we see you and appreciate you!

A banner weekend for Arcee as Trinity Large Preview officially entered the Top 5 overall models on @OpenRouter! After surpassing 80B tokens on Sunday, we are already on pace to break that record again today. As our inference volume consistently tests the upper bound of our current limits, this growth signals a compounding market demand for our frontier-class open weights as we scale far beyond our initial 1T token milestone. Stay tuned—more exciting news is on the horizon!

Grateful for such an incredible team.

We don’t want to have to choose between 2 model providers We want to choose between 1,000s of model providers