← All models

Hermes 4 14B

Nous Research · 14.8B parameters

Hermes 4 14B is Nous Research's hybrid-mode reasoning fine-tune of Qwen3-14B, a smaller sibling of the Llama-3.1-based Hermes 4 70B and 405B. Nous reports the Hermes 4 post-training corpus as roughly 5 million samples and 60 billion tokens, so no exact training-token count is recorded; extended thinking in <think> segments is enabled through the chat template's thinking flag or a reasoning system prompt. Nous documents no context length for this checkpoint, and its released config sets 40,960 maximum positions.

Benchmark scores