Nous Research

A post-training lab has no base models of its own, so its history is a list of other people's. Nous Research's is unusually legible, because OpenRouter never deletes an author page. Read theirs from the bottom and it goes: Hermes 13B and 70B on Llama 2, with two Capybara releases beside them; the Hermes 2 family on Mistral 7B, on Mixtral 8x7B in both DPO and SFT cuts, and on Yi 34B; Hermes 2 Pro and Theta on Llama 3 8B; DeepHermes 3 on Mistral 24B and on Llama 3 8B; then Hermes 3 and Hermes 4 on Llama 3.1 at 70B and 405B. It is a record of which open base was worth building on in a given quarter, kept by a lab that had to choose one every time.

The choosing stopped being visible there on 26 August 2025, when Hermes 4 70B and 405B were listed — still on the Llama 3.1 weights the router had been serving Hermes 3 from since August 2024. Four days later Nous published Hermes-4-14B on Qwen/Qwen3-14Bsource, accessed 2026-09-06. In November came Hermes 4.3 on ByteDance-Seed/Seed-OSS-36B-Basesource, accessed 2026-09-06; in December, Nomos 1, a proof-writing specialisation of Qwen/Qwen3-30B-A3B-Thinking-2507source, accessed 2026-09-06 built with Hillclimb AI, whose card scores it at 87/120 on Putnam 2025 in the Nomos reasoning harness, against 24/120 for its own base model under the same conditionssource, accessed 2026-09-06. The router lists none of the three, and the most-liked repository in the whole Hermes 4 line is among them: Hermes-4.3-36B has 299 likes for Hermes-4.3-36Bsource, accessed 2026-09-06.

So the Llama chapter is the only one the catalog carries, and it closed while it was being written. What replaced it is not merely a different vendor's weights. Hermes 4.3 is, in Nous's words, "our first production model post-trained entirely on the Psyche network"source, accessed 2026-09-06 — the distributed-training network it had announced six months earlier, on 14 May 2025, which shards a run across idle hardware and settles who did what through Solana smart contracts. Distributed training is what Nous is known for and had never put inside a shipped model; the first one that carries it is among the three the router does not list.

Psyche's own first run went further and hid better. Consilience 40B was pretrained from scratch over the internet — 40 billion parameters, "DeepSeek v3 + MLA (Dense version without MoE routers)", pretrained on 20T tokens with the DisTrO optimizersource, accessed 2026-09-06 — with no annealing pass, on the stated ground that annealing "may potentially constrain creativity and interesting behaviors", and released under dual — CC0 by default, "while also allowing it to be used under the MIT license for users who require permissive terms with attribution and warranty disclaimers"source, accessed 2026-09-06. The lab whose entire router history is other companies' bases had built one of its own and put it in the public domain. It sits on Hugging Face under PsycheFoundation, not NousResearch, so it is invisible to anyone browsing the lab's own model list.

The last thing a four-row column cannot show is that Nous now runs a catalog of its own. Nous Portal bills itself as "Nous Portal — one account for Hermes Agent, the Inference API, and Hermes Cloud. Hundreds of models, tools included, free to start."source, accessed 2026-09-06, and what its landing page renders is a router's inventory: Anthropic, OpenAI, Google, DeepSeek, Meta, Qwen, Mistral, xAI and two dozen others, priced per million tokens in and out. Counted on 6 September 2026 it held 333 distinct model ids, counted from 346 landing-page rows after excluding 13 ~-prefixed free-tier duplicate ids — and not one of them a Nous modelsource, accessed 2026-09-06. An open-weights lab whose four listed models sit in somebody else's aggregator has built an aggregator, stocked it with everybody else's models, and shelved none of its own.