𝗡𝗼𝘂𝘀 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗧𝗲𝗮𝗺 𝗥𝗲𝗹𝗲𝗮𝘀𝗲𝘀 𝗛𝗲𝗿𝗺𝗲𝘀 𝟰: 𝗔 𝗙𝗮𝗺𝗶𝗹𝘆 𝗼𝗳 𝗢𝗽𝗲𝗻-𝗪𝗲𝗶𝗴𝗵𝘁 𝗔𝗜 𝗠𝗼𝗱𝗲𝗹𝘀 𝘄𝗶𝘁𝗵 𝗛𝘆𝗯𝗿𝗶𝗱 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴

Nous Research has officially unveiled Hermes 4, a powerful new family of open-weight AI models available in 14B, 70B, and 405B parameter sizes, all built on Llama 3.1 checkpoints. What makes Hermes 4 special is that it reaches frontier-level performance using only post-training techniques — no secret data, no closed methods.

One of its standout features is hybrid reasoning. These models can switch between standard responses and explicit reasoning using special tags whenever a problem requires deeper deliberation. This gives Hermes 4 the ability to handle complex tasks more intelligently while still providing concise answers when needed.

What’s even more impressive is that Hermes 4 sets a new benchmark among open-weight models, delivering state-of-the-art performance while staying completely transparent and following a neutral alignment philosophy. In other words, it proves that cutting-edge reasoning capabilities can be developed entirely through open-source methods — a huge step forward for the AI community.

𝗗𝗮𝘁𝗮𝗙𝗼𝗿𝗴𝗲: 𝗚𝗿𝗮𝗽𝗵-𝗕𝗮𝘀𝗲𝗱 𝗦𝘆𝗻𝘁𝗵𝗲𝘁𝗶𝗰 𝗗𝗮𝘁𝗮 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻

At the heart of Hermes 4 lies DataForge, a groundbreaking system that’s changing the way training data is created. Unlike traditional methods where datasets are manually curated, DataForge takes a graph-based approach to synthetic data generation, making the entire process far more powerful and flexible.

Here’s how it works: DataForge is built on a directed acyclic graph (DAG), where each node represents a specific action defined through PDDL (Planning Domain Definition Language). Every node comes with its own preconditions, postconditions, and transformations, which together enable the automatic creation of complex data pipelines.

By leveraging pre-training seed data from sources like DCLM and FineWeb, DataForge can do some truly impressive things. For example, it can take a Wikipedia article, transform it into a rap song, and then generate instruction-answer pairs based on that transformation.

This innovative system produces around 5 million samples containing a whopping 19 billion tokens. Notably, reasoning-focused samples are intentionally token-heavy — averaging five times more tokens than regular samples — to handle detailed thinking traces of up to 16,000 tokens.

In short, DataForge is what gives Hermes 4 its edge, enabling smarter, richer, and more context-aware reasoning by revolutionizing how training data is synthesized.

𝗥𝗲𝗷𝗲𝗰𝘁𝗶𝗼𝗻 𝗦𝗮𝗺𝗽𝗹𝗶𝗻𝗴 𝗮𝘁 𝗨𝗻𝗽𝗿𝗲𝗰𝗲𝗱𝗲𝗻𝘁𝗲𝗱 𝗦𝗰𝗮𝗹𝗲

Hermes 4 relies on Atropos, Nous Research’s open-source reinforcement learning environment, to deliver its advanced reasoning capabilities. Atropos plays a crucial role by implementing rejection sampling across nearly 1,000 different task-specific verifiers, ensuring that the model learns only from high-quality reasoning trajectories spread across a variety of domains.

Some of the key verification environments used in this process include:

Answer Format Training – rewards the model for maintaining correct formatting across 150+ output formats.

Instruction Following – leverages RLVR-IFEval tasks to handle complex instructions and constraints more effectively.

Schema Adherence – ensures accurate JSON generation by validating outputs against Pydantic models.

Tool Use Training – teaches the model how to demonstrate agent-like behavior when using external tools.

Through this extensive rejection sampling process, Atropos builds a large corpus of verified reasoning trajectories, often creating multiple unique solution paths that lead to the same correct result. This approach ensures that Hermes 4 learns robust reasoning strategies rather than simply memorizing fixed templates.

In essence, Atropos acts as the quality gatekeeper for Hermes 4, shaping it into a more reliable, adaptive, and intelligent reasoning system.