Hermes 4
Nous Research's open-weight family with hybrid <think> reasoning: 14B, 70B, and 405B under Apache 2.0.
About
Hermes 4 is Nous Research's open-weight model family released August 27, 2025, in 14B, 70B, and 405B sizes. The 14B and 70B are built on Qwen3 bases and the 405B on Llama 3.1, all post-trained with Nous's DataForge synthetic-data system (about 5M samples / 19B tokens) and a hybrid reasoning design: the model can toggle between direct answers and explicit <think>...</think> reasoning traces up to 16K tokens long. The release was paired with a 94-page technical report (arXiv:2508.18255) and a stated neutral-alignment philosophy with minimal refusal training.
Hermes is the model arm of a lab that raised a $50M Paradigm-led Series A at a $1B valuation in 2025 and runs the Psyche decentralized training network; Hermes 4.3 (36B, December 2025) was the first model trained in full on Psyche. Nous positions Hermes 4 for agentic use, function calling, and structured outputs, and the family has become a durable default for self-hosted agent stacks. No canonical independent benchmark table was found, so capability claims rest on the technical report and community evals.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No canonical independent benchmark table found; Nous reports frontier-level open-weight performance with strong math results in the 94-page technical report (arXiv:2508.18255) — treat as provider-reported
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: public release Aug 27, 2025 across 14B/70B/405B with Apache-2.0 weights on Hugging Face; 94-page technical report with data methodology (arXiv:2508.18255); lab funded ($50M Series A, Paradigm) with the live Psyche training network behind follow-ups. Open weights. Release facts corroborated across MarkTechPost, the Nous timeline, and the HF model pages. Strengths: transparent post-training story, permissive license, real agentic adoption. Docked for: (1) no independent benchmark table found — performance claims rest on the vendor's own report; (2) minimal-refusal alignment posture is a deliberate choice that narrows some enterprise use.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Nous Research release 2025-08-27; 14B/70B (Qwen3) + 405B (Llama 3.1) on Hugging Face
- · 94-page technical report arXiv:2508.18255; hybrid <think> reasoning
- · Lab funded ($50M Paradigm Series A); Hermes 4.3 trained on Psyche Dec 2025
- · No independent benchmark table found; minimal-refusal alignment narrows some enterprise use