Hermes 4.3 36B — Tools and Thinking

A correctly-configured Ollama upload of Nous Research’s Hermes 4.3 36B with both tools and thinking capabilities properly exposed.

What’s fixed

Existing community uploads of Hermes 4.3 36B advertise only “Text” capability on Ollama, causing 400 errors when used with agent frameworks that pass tool definitions. The model itself is fully tool-trained — the gap was in the modelfile template, which didn’t include the conditional structures Ollama’s parser uses to detect tool support.

This upload fixes:

• Tools capability properly advertised via correct template structure
• Thinking capability properly advertised with system-prompt switching on the think parameter
• Chat template aligned with Nous’s canonical Jinja template (Llama-3 ChatML headers, not OpenAI ChatML as some uploads use)
• Stop sequences matched to what the model actually generates
• Default context set to 32K instead of the model’s 524K maximum (which otherwise consumes most VRAM/unified memory in KV cache)

Usage

Standard chat: ollama run steelpuddles/hermes-4.3-36b:tools-thinking

With thinking enabled (default for capable models): curl http://localhost:11434/api/chat -d ‘{ “model”: “steelpuddles/hermes-4.3-36b:tools-thinking”, “messages”: [{“role”: “user”, “content”: “your prompt”}], “think”: true }’

With tools (works with Hermes Agent and other OpenAI-compatible agent frameworks): curl http://localhost:11434/api/chat -d ‘{ “model”: “steelpuddles/hermes-4.3-36b:tools-thinking”, “messages”: [{“role”: “user”, “content”: “What is the weather?”}], “tools”: [{“type”: “function”, “function”: {…}}] }’

License

Apache 2.0 throughout the chain: - Base model: ByteDance Seed-OSS-36B (Apache 2.0) - Fine-tune: NousResearch Hermes 4.3 36B (Apache 2.0) - Quantization: HammerAI’s GGUF build - Modelfile: this upload

Commercial use is permitted under Apache 2.0 terms.

Credits

• ByteDance Seed Team for the Seed-OSS-36B base model
• Nous Research for Hermes 4.3 fine-tuning and the canonical chat template
• HammerAI for the GGUF quantization
• Modelfile fixes by Steel Puddles (steelpuddles.substack.com)

Notes

The model’s thinking activation switches the system prompt to Nous’s deep-thinking prompt when think=true is passed. When think=false or when the field is omitted (depending on Ollama CLI version), the standard Hermes system prompt is used.

For agent framework users hitting “Hermes models are not agentic” warnings: that warning is pattern-matched on the model name string and does not reflect actual capabilities. This upload supports tool calling fully.