Nous Research drops API for unrestricted Hermes 3 models

If you have ever had Claude or GPT flat-out refuse a completely harmless query because an over-sensitive safety filter tripped, you know the frustration. The obvious escape hatch has always been running open-source models yourself, but keeping a beefy GPU rig humming just to test a prompt is a massive headache.

Nous Research, the New York-based collective behind the popular uncensored Hermes series, just launched its own Inference API to cut out that friction entirely.

Drop-in OpenAI compatibility and two flagship models

The initial rollout puts two of the team's best models behind a cloud endpoint. First up is Hermes 3 Llama 70B, their powerhouse general-purpose model built on Meta's Llama 3.1 architecture. The second is DeepHermes-3 8B Preview, a specialized reasoning model dropped last month that lets you toggle chain-of-thought thinking via system prompts.

Smartly, Nous Research made the endpoints follow OpenAI's exact API design for completions and chat completions. If your codebase already talks to GPT-4, pointing it at Hermes is practically a one-line config change.

Because Nous does not sit on an endless pile of GPUs, access is currently managed through a first-come, first-serve waitlist. Every new account gets $5 in free credits to kick the tires.

Going from GitHub releases to a real business

Since kicking off in 2023, Nous Research has openly framed itself as the counterweight to sanitized corporate AI. The collective has shipped 29 artifacts since July 2023, spanning open weights, research papers, datasets, and code. Four months ago, they dipped their toes into consumer tooling with the web-based Nous Chat.

While Nous champions "unrestricted" models aimed at maximizing individual agency, the models are not totally lawless and still retain baseline guardrails against dangerous harm. But compared to the rigid guardrails of Big Tech, Hermes offers developers far more creative breathing room.

Launching a paid inference endpoint marks a pivotal moment for the collective: finding a way to pay staggering server bills without selling out the open-source ethos that made them famous.

Why it matters

For developers building creative writing tools, roleplay engines, or experimental research pipelines, this API eliminates the painful trade-off between strict corporate guardrails and painful self-hosting setup times. Having access to genuine reasoning and high-parameter open models without wrestling with local vLLM instances is an immediate win.

We think this is a clever survival strategy. Open-source AI projects routinely starve to death because giving away weights does not pay for compute. By charging for hosted convenience while keeping the weights free, Nous creates a viable revenue loop. Copying OpenAI's API syntax is an aggressive, pragmatic touch that lowers switching costs to near zero.

The real test will be reliability. Nous is candidly resource-constrained compared to cloud hyperscalers. The waitlist buys them time, but if their inference clusters choke during traffic spikes, developers will hesitate to run production workloads on top of them.