Nous Research Unleashes Hermes 4: Open-Source AI Models

Nous Research, a pioneering group in open-source and decentralized AI, has launched Hermes 4, their most advanced family of large language models to date.

Built for transparency and user steerability, Hermes 4 stands out for its reasoning capabilities, performance, and minimal content restrictions.

Key Technical Features

Hybrid Reasoning Architecture: Hermes 4 introduces “hybrid reasoning” capabilities, allowing users to toggle between quick responses and thorough, step-by-step cognitive processes.

When activated, models articulate their internal reasoning within special <think> tags before delivering final responses, similar to OpenAI’s o1 reasoning models but with complete transparency regarding the AI’s thought process.

Function Calling and Tool Use: Hermes 4 supports advanced function calling within single assistant turns, produced after reasoning processes.

The system handles tool definitions through system prompts or direct message fields, with automatic parsing built into VLLM and SGLang frameworks.

Advanced Training Infrastructure: Hermes 4 leverages two revolutionary training systems: DataForge for synthetic data generation and Atropos for reinforcement learning.

DataForge operates through directed graphs, transforming basic pre-training data into complex instruction-following examples through “random walks”.

For instance, it can convert a Wikipedia article into a rap song before generating related questions and answers.

Atropos functions as specialized training environments where AI models practice specific skills mathematics, coding, tool utilization, and creative writing receiving feedback only for correct responses.

This “rejection sampling” method ensures only verified, high-quality responses enter training data.

Model Specifications and Performance

Available Model Sizes: Hermes 4 comes in three configurations: 14B, 70B, and 405B parameters, all built on Llama 3.1 architecture (except 14B using Qwen3 14B checkpoint).

Each model offers hybrid reasoning capabilities with toggleable thinking modes.

Benchmark Performance

Hermes 4’s performance achievements are remarkable:

• 405B model: 96.3% on MATH-500 benchmark in reasoning mode, 81.9% on AIME’24 mathematics competition.
• 70B model: 95.6% on MATH-500, 73.5% on AIME’24.
• 14B model: 92.6% on MATH-500, 52.7% on AIME’24.
• Most significantly, Hermes 4 achieved 57.1% on RefusalBench in reasoning mode, vastly outperforming GPT-4o (17.67%) and Claude Sonnet 4 (17%).

Availability and Access Options

All model weights are freely available on Hugging Face following Nous Research’s open-source philosophy.

Interactive Platforms

Users can access Hermes 4 through multiple channels:

• Nous Chat: Revamped interface featuring parallel interactions and memory systems.
• API Access: Available through OpenRouter, Nebius, and Luminal.
• Direct Download: Complete model weights on Hugging Face platform.

Pricing Structure

Hermes 4 maintains cost-effective pricing compared to proprietary alternatives:

• Input tokens: $0.000093 per 1,000 tokens.
• Output tokens: $0.000373 per 1,000 tokens.
• Context length: 131,072 tokens.

This pricing significantly undercuts competitors like Claude 3.5 Sonnet ($0.003/$0.015) and GPT-4o while providing comparable or superior performance.

News Gist

Nous Research has launched Hermes 4, a family of open-weight hybrid reasoning AI models (14B, 70B, 405B).

With benchmark-leading performance, hybrid reasoning mode, and transparent training, Hermes 4 challenges Big Tech dominance while promoting openness, affordability, and user control.

FAQs

Q1. What is Hermes 4?

Hermes 4 is a family of open-weight hybrid reasoning AI models (14B, 70B, 405B) released by Nous Research in August 2025.

Q2. What makes Hermes 4 unique?

It features a hybrid reasoning mode, benchmark-topping math performance, fewer restrictions on responses, and complete transparency in training methods and datasets.

Q3. How does Hermes 4 perform in benchmarks?

The 405B version achieved 96.3% on MATH-500 and 81.9% on AIME’24, rivaling or surpassing top closed-source AI models.

Q4. Where can Hermes 4 be accessed?

It is available on Hugging Face, Nous Chat, and OpenRouter, with both free model weights and paid inference services.

Q5. What is the pricing for Hermes 4 usage?

On OpenRouter, pricing is about $0.20 per million input tokens and $0.80 per million output tokens for the 405B variant.

Q6. Why is Hermes 4 significant?

Hermes 4 challenges Big Tech by showing that open-source AI can be powerful, affordable, transparent, and user-controlled, empowering developers and researchers worldwide.