Hermes 4.3 36B — Tools + Thinking (Q8_0)

A properly-configured Ollama packaging of Nous Research’s Hermes 4.3 36B, built on the Q8_0 GGUF with the correct Llama-3 chat template and verified tools + thinking capabilities.

Maintained by Monomyth Development.

What this fixes

Hermes 4.3 36B is fully tool-trained, but many community Ollama/GGUF uploads advertise only completion (“Text”) capability — agent frameworks that pass a tools array then error or silently lose tool calling. The cause is the Modelfile template, not the weights: uploads frequently ship a ChatML template (<|im_start|> / <|im_end|>), which is the Hermes-4 14B format. Hermes 4.3 36B is Llama-3 (<|start_header_id|> / <|eot_id|>).

This build applies the correct Llama-3 template — adapted from steelpuddles/hermes-4.3-36B:thinking-tools, who did the original template work — with the conditional structures Ollama’s parser reads to detect capabilities:

• Tools capability via the .Tools template branch
• Thinking capability, switched on the native think request parameter
• Llama-3 chat template (not ChatML), with <tool_call> / <tool_response> framing
• Stop sequences matched to model output (<|eot_id|>, <|end_of_text|>)
• Default context raised to 32K (the stock 4096 truncates tool definitions)
• Q8_0 quant — prioritizes quality over footprint

Confirm after pulling:

ollama show MonomythDevelopment/hermes-4.3-36b-tools
# Capabilities: completion, tools, thinking

Quick start

ollama pull MonomythDevelopment/hermes-4.3-36b-tools
ollama run  MonomythDevelopment/hermes-4.3-36b-tools "What's 2+2?"

Thinking (per-request toggle)

Thinking is a reasoning mode (the model emits <think>…</think> before answering), mapped to Ollama’s native think field — orthogonal to tool calling, controlled independently. Default it off for agent loops (no <think> blocks to strip from tool-call output) and opt in where deliberation helps.

curl http://localhost:11434/api/chat -d '{
  "model": "MonomythDevelopment/hermes-4.3-36b-tools",
  "messages": [{"role": "user", "content": "Prove that sqrt(2) is irrational."}],
  "think": true,
  "stream": false
}'

Tool calling

Tools are declared as OpenAI-style JSON schemas; the model emits <tool_call>{"name": …, "arguments": {…}}</tool_call>; results return in <tool_response>…</tool_response>. Any OpenAI-compatible client that sends a tools array works — verify by confirming message.tool_calls is populated.

curl http://localhost:11434/api/chat -d '{
  "model": "MonomythDevelopment/hermes-4.3-36b-tools",
  "messages": [{"role": "user", "content": "What is the weather in Paris?"}],
  "tools": [{
    "type": "function",
    "function": {
      "name": "get_weather",
      "description": "Get current weather for a city",
      "parameters": {
        "type": "object",
        "properties": {"city": {"type": "string"}},
        "required": ["city"]
      }
    }
  }],
  "stream": false
}'

Runtime caveat: seed_oss tool-call parsing in llama.cpp/Ollama is still maturing — a correct template can’t fix an immature parser underneath it. If tool calls parse unreliably, update to the latest Ollama, or serve with vLLM (--enable-auto-tool-choice --tool-call-parser hermes). Verify on your own build before relying on it.

Parameters

License

Apache 2.0 throughout — commercial use permitted. Copyright 2026 Monomyth Development.

Credits

This is a packaging of others’ work; it adds no weights of its own.

• ByteDance Seed Team — base model (Seed-OSS-36B-Base)
• Nous Research — fine-tune and Q8_0 GGUF (Hermes-4.3-36B, GGUF)
• Steel Puddles — original Llama-3 tools+thinking template (steelpuddles/hermes-4.3-36B)

Source repo & full attribution: https://github.com/MonomythDevelopment/ollama-hermes-4.3-36b-tools