Hermes 3 - 2026 Pricing, Features, Reviews & Alternatives

Hermes 3 is a series of open-weight instruct and tool-use language models fine-tuned from Llama 3.1 across three parameter scales: 8B, 70B, and 405B. Trained on approximately 390 million tokens of primarily synthetically generated instruction data, the models target high steerability, precise instruction-following, multi-turn conversation, agentic function-calling, and creative tasks including complex roleplaying. Weights are available on HuggingFace for self-hosted deployment, with inference also accessible through the Nous Portal and compatible third-party endpoints such as OpenRouter.

Core capabilities include a 128K token context window for long-term context retention across extended multi-turn conversations, and agentic function-calling via the Hermes Function Calling standard using JSON schema-defined tool definitions. Retrieval-augmented generation (RAG) is supported with structured citation tags, enabling grounded responses from external data sources. Structured output is produced through a defined set of XML tokens including SCRATCHPAD, REASONING, INNER_MONOLOGUE, PLAN, EXECUTION, REFLECTION, THINKING, SOLUTION, EXPLANATION, and UNIT_TEST, making multi-step reasoning transparent and interpretable. The 405B variant exhibits distinct behavior sensitivity to empty system prompts, offering additional control over model persona and response style.

Hermes 3 targets researchers, open-source AI developers, and technical practitioners building agentic pipelines, tool-use applications, RAG systems, and creative or code generation workflows. Training data was filtered for quality through token length thresholds, removal of refusals, and prioritization of outputs from high-performing models. The result is a model suite that delivers benchmark-competitive performance among open-weight alternatives while exposing deeper reasoning, planning, and creative capabilities not typically available in closed commercial models.