Hermes 3
A family of open-source large language models from Nous Research fine-tuned on Llama 3.1, designed for agent tasks with native tool calling, structured JSON generation and multi-turn reasoning, available in 8B, 70B and 405B parameter sizes.
Timeline (last 90 days)
Hermes 3 can be seamlessly integrated into orchestration frameworks such as LangChain and CrewAI through standard connectors; for example, in LangChain you only need to point ChatOpenAI's base_url to the local endpoint, set model to hermes3:8b, and then bind tools via create_tool_calling_agent.
Hermes 3 is an open-source large model fine-tuned by Nous Research based on Llama 3.1, designed specifically for Agent tasks, supporting multi-turn reasoning, structured JSON generation, and native tool calling.
Hermes 3 uses the structured tags <tool_call>...</tool_call> to format function calls, thereby reducing the probability of JSON syntax corruption.
Hermes 3 is offered in three parameter sizes—8B, 70B, and 405B—of which 8B is suitable for a local Mac or a single consumer-grade GPU, while 70B/405B are suitable for enterprise clusters.
vLLM's --tool-call-parser hermes parameter and Ollama's built-in parser can convert the XML-like <tool_call> tags generated by Hermes 3 into the OpenAI-compatible tool_calls schema.
When building the Agent loop for Hermes 3, it is recommended to set max_iterations=5 and temperature=0.2, and when tool execution throws an exception, to catch it and return it to the model as {"error": ...} so that the model has a chance to self-correct.
When a tool returns 404 Not Found or a SQL Syntax Error, Hermes has the intrinsic ability to analyze the traceback and generate a corrected call.
The Hermes series has currently iterated to Hermes 3, with the base model typically selected from Meta's Llama series or the Mistral series.
All Facts (8)
Hermes 3 can be seamlessly integrated into orchestration frameworks such as LangChain and CrewAI through standard connectors; for example, in LangChain you only need to point ChatOpenAI's base_url to the local endpoint, set model to hermes3:8b, and then bind tools via create_tool_calling_agent.
50%UnverifiedHermes 3 is an open-source large model fine-tuned by Nous Research based on Llama 3.1, designed specifically for Agent tasks, supporting multi-turn reasoning, structured JSON generation, and native tool calling.
50%UnverifiedHermes 3 uses the structured tags <tool_call>...</tool_call> to format function calls, thereby reducing the probability of JSON syntax corruption.
50%UnverifiedHermes 3 is offered in three parameter sizes—8B, 70B, and 405B—of which 8B is suitable for a local Mac or a single consumer-grade GPU, while 70B/405B are suitable for enterprise clusters.
50%UnverifiedvLLM's --tool-call-parser hermes parameter and Ollama's built-in parser can convert the XML-like <tool_call> tags generated by Hermes 3 into the OpenAI-compatible tool_calls schema.
50%UnverifiedWhen building the Agent loop for Hermes 3, it is recommended to set max_iterations=5 and temperature=0.2, and when tool execution throws an exception, to catch it and return it to the model as {"error": ...} so that the model has a chance to self-correct.
50%UnverifiedWhen a tool returns 404 Not Found or a SQL Syntax Error, Hermes has the intrinsic ability to analyze the traceback and generate a corrected call.
50%UnverifiedThe Hermes series has currently iterated to Hermes 3, with the base model typically selected from Meta's Llama series or the Mistral series.
50%