Model details
Hermes-4-70B
Hermes-4-70B is a hybrid-mode reasoning model built upon the Llama-3.1-70B architecture, designed to balance direct responses with deep analytical deliberation. The model features a flexible reasoning system that allows users to toggle explicit thinking traces, providing a transparent look at its logic when tackling complex problems. It is engineered to excel in technical domains such as mathematics, coding, and STEM, while simultaneously maintaining high proficiency in creative writing and subjective tasks. By prioritizing format-faithful outputs and schema adherence, the model serves as a robust tool for developers requiring reliable, structured data generation alongside general-purpose assistant capabilities.
The model benefits from an extensive post-training lineage, utilizing a massive, synthesized corpus of approximately 60 billion tokens across 5 million samples. This training approach emphasizes verified reasoning traces, which significantly enhances the model's logical consistency and instruction-following accuracy compared to its predecessors. Beyond its reasoning capabilities, the model is optimized for improved steerability and reduced refusal rates, making it highly adaptable to specific user values and complex roleplay scenarios. With its ability to repair malformed objects and adhere strictly to provided schemas, it is well-positioned for integration into workflows that demand both creative flexibility and rigorous technical precision.
Quick Info
Powered by- Provider
- Nebius Token Factory
- Model key
- NousResearch/Hermes-4-70B
- Release date
- Jan 30, 2026
- Last updated
- Feb 4, 2026
- Knowledge cutoff
- 2025-11
- Input modalities
- Output modalities
- Capabilities
Cost
- Input token cost
- $0.13
- Output token cost
- $0.40
Limits
- Input tokens
- 120,000 tokens
- Output tokens
- 8,192 tokens
- Context window
- 128,000 tokens
Latest news about Hermes-4-70B
No articles yet. Fetch the latest news to show it here.