Hermes 4: Open-Source AI That Says “Yes,” Not “I’m Sorry”
Nous Research has just released Hermes 4, an open-weight language model touted as a GPT-4o-caliber competitor, but with a bold twist: it skips built-in censorship and defers decisions entirely to the user. It's “what if AI actually answered?”
The Hermes 4 family spans three sizes—14, 70 and a flagship 405 billion parameters—each trained on a cocktail of synthetic data and reinforcement-learning loops designed to sharpen reasoning, coding and logic. The results are difficult to dismiss: a staggering 96.3% on the notoriously brutal MATH-500 benchmark and 61.3% on LiveCodeBench, placing it shoulder-to-shoulder with the strongest commercial models.
Where Nous departs most radically from the establishment is in philosophy. Hermes 4 doesn’t apologise, doesn’t dodge, and doesn’t refuse. On RefusalBench, a new test measuring how often an AI declines to answer, the 405B model scored 57.1%, compared with GPT-4o’s meagre 17.7%. To its makers, this is proof of “user sovereignty.” To its critics, it’s a Pandora’s box.
The mechanics are as ambitious as the ethos. A new data pipeline, dubbed DataForge, spins diverse synthetic training samples from vast graph structures, while the Atropos reinforcement environment forces models to practise problems hundreds of thousands of times, admitting only verified solutions into the dataset. To prevent runaway rambling, Hermes was deliberately trained to cut off reasoning at around 30,000 tokens, balancing transparency with usability.
That transparency is embodied in its hybrid reasoning system: users can choose between a snappy reply or a fully annotated step-by-step process, exposed in <think>…</think> tags. In theory, that means you can not only get the answer but also watch the machine think—an idea equal parts fascinating and unnerving.
On forums from Reddit to Hacker News, Hermes 4 has been hailed as “the most open release yet,” with some calling its edgy system prompts “straight from a ’90s anime.” Others see danger in the absence of built-in guardrails. Either way, Nous has ensured Hermes 4 won’t be ignored.
In an AI world obsessed with safety disclaimers and curated politeness, Hermes 4 stands out as something brash, raw and unapologetically open. Whether it proves to be the vanguard of a freer future or a cautionary tale of too much, too soon, may define how the next generation of artificial intelligence is built—and who controls it.
