A moe model you say? How kawaii is it? uwu

Capitalization makes a surprising amount of difference here...

Meccha at present, but it may reach sugoi levels with fine-tuning.

I hate that I laughed at this. Thanks ;)

Trinity Nano Preview: 6B parameter MoE (1B active, ~800M non-embedding), 56 layers, 128 experts with 8 active per token

Trinity Mini: 26B parameter MoE (3B active), fully post-trained reasoning model

They did pretraining on their own and are still training the large version on 2048 B300 GPUs

Looks like a less good version of qwen 30b3a which makes sense bc it is slightly smaller. If they can keep that effiency going into the large one it'll be sick.

Trinity Large [will be] a 420B parameter model with 13B active parameters. Just perfect for a large Ram pool @ q4.

> Trinity Large is currently training on 2048 B300 GPUs and will arrive in January 2026.

How long does the training take?

Couple days or weeks usually. No one is doing 9 month training runs

Excited to put this through its paces. It seems most directly comparable to GPT-OSS-20B. Comparing their numbers on the Together API: Trinity Mini is slightly less expensive ($0.045/$0.15 v $0.05/$0.20) and seems to have better latency and throughput numbers.

Moe ≠ MoE

?

The HN title uses incorrect capitalization.

I was eagerly waiting for the Larry and Curly models.

^_-

Interesting. Always glad to see more open weight models.

I do appreciate that they openly acknowledge the areas where they followed DeepSeek's research. I wouldn't consider that a given for a US company.

Anyone tried these as a coding model yet?