Hermes in one paragraph
Hermes is a family of large language models from Nous Research that you can download and run yourself. The weights are on Hugging Face, the models do function calling and structured JSON output natively, and a hybrid reasoning mode turns deliberation on and off per request. As of September 2026 the line is Hermes 4 — released 26 August 2025 in 14B, 70B and 405B sizes — plus Hermes 4.3 36B, published December 2025 and post-trained on Nous Research's decentralized Psyche network.
The interesting part for a business is not the benchmark table. It is the ownership model. A model whose weights sit on your disk cannot be deprecated out from under you, cannot reprice mid-quarter, and does not require customer data to leave your network — and it carries a bill of its own.
What "open weights" means, and who builds Hermes
Open weights means the trained parameters are downloadable. You get the file, so you can run it on your own hardware, quantize it, fine-tune it, serve it behind your own API gateway, and keep running that exact version in five years.
It is not open source in the software sense — no training data, no training code. You can audit the model's behaviour, not its provenance. It is also not a closed API model: there, the vendor absorbs the hardware and upgrades while your prompts cross their infrastructure, your unit cost stays per-token forever, and your tested version retires on their schedule. Open weights inverts every one of those.
Nous Research is an open-source AI lab founded in 2023 by Jeffrey Quesnelle, Karan Malhotra, Ryan Teknium and Shivani Mitra. It raised a $50 million Series A led by Paradigm in April 2025, and TechCrunch reported in July 2026 that it was finalising a further round at a $1.5 billion valuation — in negotiation then, not closed. Their position, visible throughout Hermes: models should be open, downloadable and steerable by whoever deploys them.
The Hermes 4 line: sizes, base models and context lengths
Nous post-trains other organisations' base models: it does the instruction tuning, reasoning training and alignment, and the pretrained foundation comes from Meta, Qwen or ByteDance. That determines both the licence you inherit and the context window you get.
Hermes 4 405B is a hybrid-mode reasoning model based on Llama-3.1-405B, 131,072-token context. Hermes 4 70B is based on Llama-3.1-70B, also 131,072. Hermes 4 14B is based on Qwen3-14B at 40,960. Hermes 4.3 36B, the newest, is based on ByteDance's Seed-OSS-36B and supports 524,288 tokens — by a wide margin the longest in the family. Those figures come from each model's own config on Hugging Face; re-check them rather than trusting a blog post a year from now.
Nous's published numbers put Hermes 4.3 36B at roughly the level of Hermes 4 70B — MMLU 87.7 against 88.4, AIME 25 at 69.3 against 67.4, GPQA Diamond 65.5 against 66.1 — at about half the parameters. The same table shows a real weakness: SimpleQA, measuring broad factual recall, scores 6.0 for the 36B against 17.9 for the 70B. For open-domain factual Q&A with no retrieval layer, that gap will surface in production.
How Hermes 4 differs from Hermes 3
Nous grew the post-training corpus from roughly 1M samples and 1.2B tokens to approximately 5M samples and 60B tokens, weighted toward verified reasoning traces. The structural additions are hybrid-mode reasoning and much stronger schema adherence: Hermes 4 is trained to produce valid JSON against a supplied schema and to repair malformed objects.
What hybrid reasoning means in practice
One model, two modes, chosen per request. In reasoning mode it emits an explicit internal monologue inside <think> tags before answering; otherwise it responds directly. You switch with a chat-template flag (thinking=True) or the deep-thinking system prompt on the model card. Reasoning tokens are still tokens, so deliberating on contract analysis and answering directly on a routing classifier from one endpoint means you run one model, not two.
Does Hermes support tool calling and structured outputs?
Yes, trained in rather than bolted on. Hermes emits function calls inside <tool_call> tags within a single assistant turn, after its reasoning; those tags are added tokens, so they parse cheaply while streaming. Pass tool definitions in the messages "tools" field and the chat template builds the system prompt. Both serving backends ship parsers: in vLLM set the tool parser to hermes, in SGLang use qwen25.
The licence question, answered properly
Here are the actual licence tags from the model cards, checked this month. Hermes 4 14B and Hermes 4.3 36B are Apache 2.0 — permissive, explicit patent grant, no user-count threshold, no naming or attribution obligations on your product.
Hermes 4 70B and 405B are not Apache 2.0. Both carry the llama3 tag, inheriting the Llama 3.1 Community License from their Meta base model. It permits commercial use royalty-free, with conditions: display "Built with Llama" where you distribute the materials or products containing them; any model you train on Llama materials must be named beginning with "Llama"; Meta's Acceptable Use Policy applies; and organisations above 700 million monthly active users on the Llama 3.1 release date need a separate Meta licence.
For nearly every small and mid-sized business the 700-million clause is irrelevant and attribution is a five-minute task. But it is a bespoke community licence, not Apache or MIT, and enterprise procurement teams read licence files.
Psyche: why Nous trains models on a distributed network
Psyche is Nous Research's decentralized training network. Rather than one datacentre of co-located accelerators, it coordinates training across nodes spread over the open internet using an optimizer called DisTrO that cuts how much data nodes exchange between steps, with coordination secured by consensus on the Solana blockchain.
Hermes 4.3 36B was post-trained entirely on the Psyche network — the first Hermes model to be. Nous reported 144,000 tokens per second across 24 nodes and trained a centralized control version; their comparison shows the Psyche-trained model matching or slightly beating it on most downstream evaluations. You never touch Psyche as a user, but it signals that capable models can be trained without one organisation owning the whole compute stack.
The real business case for open-weight models
Four arguments hold up. Most of the others are rationalisations.
Data residency. When inference runs on hardware you control, prompts containing client records, medical notes, privileged communications or payroll data never cross a third-party boundary. That is a fact about your network topology, not a policy promise. It is the most common reason our clients self-host, covered in our guide to self-hosted AI for business.
Flat cost past a volume threshold. Hosted API spend scales linearly with usage; a self-hosted deployment costs roughly the same at a thousand requests a day or fifty thousand, until you saturate the GPU. That flips the economics past a certain volume — but only past it.
No deprecation. You tune prompts and evaluations against a version; eighteen months later a hosted vendor retires it and your calibrated outputs shift. With weights on disk, the model you validated is the model you keep running.
Offline operation. Secure facilities, some clinical and defence contexts, field deployments, factory floors. A downloadable model is the only thing that works with no internet egress.
The honest costs nobody puts on the landing page
"Free model" means free weights, not free inference, and that gap is where most self-hosting projects quietly fail.
The hardware arithmetic is unforgiving. Weights consume roughly two bytes per parameter at bf16, one byte at 8-bit, about half a byte at 4-bit — so a 36B model needs on the order of 72GB, 36GB or 18GB of VRAM just to hold them, before any KV cache, which grows with context length and concurrency. That is why a half-million-token window is expensive to actually use, and why a 405B model at full precision is a multi-node deployment. Our local LLM cost and hardware guide works through real configurations.
Then the recurring work with no line item until you are doing it. Someone keeps vLLM, SGLang or llama.cpp patched. Someone builds an evaluation set for your actual tasks and re-runs it whenever anything changes. Someone is reachable when the endpoint dies at 2am. Hosting is the visible cost; the operational cost is usually larger. If nobody wants to own that, a hosted API is the correct engineering decision.
Steerability and reduced refusals: a governance question, not a selling point
Hermes is often described in shorthand as "uncensored," which is not useful to a business. The plain version: Nous trains explicitly for steerability and lower refusal rates, and publishes its own benchmark, RefusalBench, measuring how often a model answers prompts other models decline. On Nous's figures, Hermes 4.3 36B answers about 74.6% of RefusalBench prompts against 59.5% for Hermes 4 70B, and Nous states Hermes 4 achieves SOTA on RefusalBench across popular closed and open models. That is their own benchmark scoring their own design goal — intent, not an independent finding.
In deployment it means the model follows your system prompt more faithfully and refuses less, which for real business work is often an advantage: consumer-tuned models decline plenty of ordinary professional tasks, from summarising a hostile legal filing to drafting a firm collections letter.
The corollary is non-negotiable. Deploy a model with weak built-in refusal behaviour and you are the safety layer — nobody moderates on your behalf, because there is no vendor in the request path. You own input filtering, prompt-injection defence, output moderation before anything reaches a customer, an acceptable-use policy, conversation logging and access control on the raw endpoint. Choosing a steerable model and skipping the guardrails is a decision to ship without them.
Where Hermes fits against other open-weight and hosted models
Hermes is not the largest or most-benchmarked open-weight family, and Nous does not claim it is. Its distinguishing traits are aggressive steerability, per-request hybrid reasoning, and tool use and structured outputs trained in rather than added afterwards.
Against the other open-weight families — Llama, Qwen, DeepSeek, Mistral, gpt-oss — base capability at a given size is broadly competitive; the real differentiators are licence terms, refusal behaviour, tool-calling ergonomics and which sizes exist. Benchmark two or three on your own tasks; no leaderboard will tell you which one handles your document format. Our roundup of the best open-source LLMs for business covers the wider field. Against a hosted frontier model, be realistic: on the hardest reasoning and broad factual recall, the leading closed models are still ahead.
So the outcome is usually not either-or. Route high-volume, sensitive, well-defined work to a self-hosted model and send the genuinely hard requests to a frontier API. That hybrid is the default architecture we build, and it is why we work bring-your-own-key — the client holds the provider accounts and API keys, so changing which model handles which route is a config change.
Deployment paths, lightest to heaviest
Most businesses should start further up this list than they expect to.
1. A hosted inference provider (start here)
Someone else runs the GPUs and you call an API, but the model is still open-weight, so you can take it elsewhere later. Nous serves its own models through the Nous Portal and a chat interface, and Chutes is listed as a provider on the model cards. Checking OpenRouter's catalogue on 1 September 2026, Hermes 4 70B was listed at $0.13 per million input tokens and $0.40 per million output, Hermes 4 405B at $1.00 and $3.00, both at 131,072 context; Hermes 4.3 36B was not listed there at the time.
This gives you zero infrastructure and full portability, but not data residency: your prompts go to a third party like any hosted API. Use it to find out whether Hermes is good enough at your task before spending a cent on hardware.
2. Your own cloud GPU
Rent a GPU instance and serve the weights with vLLM or SGLang behind your own endpoint inside your own VPC, with prefix caching on. You control the network boundary and the model version, and a per-token bill becomes an hourly one. The trap is idle time: an instance billed around the clock but used during business hours is expensive per useful token.
3. On-premise hardware
The only option for true air-gapped operation, and past a high enough sustained volume the cheapest per token — but a capital purchase depreciating against a fast-moving hardware curve. Do it when compliance leaves no alternative, or after renting hardware long enough to know what you need. GGUF quantized builds of Hermes 4.3 36B are published and the model is on LM Studio, with a Docker Model Runner path on the cards, so a single-workstation deployment is realistic in a way a 405B model never will be.
What else Nous Research makes
Hermes Agent is an MIT-licensed, Python-based autonomous agent runtime released in February 2026 — a separate product from the models, and easy to confuse with them. It runs on your own server or locally, has a CLI plus chat-platform front ends, connects to MCP servers, schedules unattended jobs on a built-in cron, spawns isolated subagents, and refines its own skills from usage. It is provider-agnostic: Nous Portal, OpenRouter, OpenAI, Anthropic or a custom endpoint. TechCrunch reported roughly 214,000 GitHub stars in July 2026 — check the repo for the current figure.
NousCoder-14B (January 2026) is a competitive-programming model on a Qwen3-14B base and Nomos 1 (December 2025) is a 30B mathematics model. Alongside them sit the older Hermes 3 and DeepHermes 3 lines and Consilience 40B, from the Psyche pretraining run.
Which industries this matters for — and who should skip it
Self-hosting earns its cost when a specific regulation, contract or client expectation makes sending data to a third party a problem — not when it merely feels safer.
Law firms are the clearest case: privilege, client confidentiality obligations and outside-counsel guidelines naming approved subprocessors. We covered the specifics in our guide to self-hosted LLMs for law firms, and the same reasoning maps onto healthcare providers handling PHI under HIPAA, accounting practices holding client financials, and any business under a contract restricting where data may be processed. Add high-volume document processing where per-token costs compound, and any operation that must continue when the internet does not.
Everyone else should probably skip it, and that is not a hedge. Under a few million tokens a month, a hosted API costs less than a day of engineering time per month. With no in-house technical owner, self-hosting creates a system nobody can fix. If your data is neither regulated nor contractually restricted, network isolation buys a feeling rather than a compliance outcome. Use a hosted API but hold your own provider API keys, and you get much of the control for none of the operational cost.
How to evaluate Hermes for your business in a week
Write down the actual task and assemble twenty to fifty real examples with the outputs you would accept. Run them against Hermes through a hosted provider and against whatever model you use today. If Hermes holds up, then price the deployment: GPU hours or hardware, plus ops time. Most self-hosting projects fail because someone priced hardware first and found out later that the model was never good enough.
This is the work we do at AutomateNexus. We are model-agnostic and BYOK by default: you hold the provider accounts and keys, model usage typically runs $30 to $150 a month paid directly to the provider with no markup from us, and a model swap is a config change. Builds start at $7,500, a typical build runs about 30 days, and an MVP is usually 4 to 8 weeks. If you want a written workflow audit and prioritised roadmap first, the paid strategy audit is $2,500.
If you are earlier than that, take the free automation audit — a three-minute questionnaire returning an automation health score, what your manual work costs you annually, and ranked quick wins. No call, no cost.
Frequently asked questions
Can I use Hermes commercially?
Yes, but check the variant. Hermes 4 14B and Hermes 4.3 36B are Apache 2.0: commercial use, no user-count threshold, no naming obligations. Hermes 4 70B and 405B carry the Llama 3.1 Community License from their Meta base models — royalty-free commercial use, but you must display "Built with Llama", any model you train on those materials must be named starting with "Llama", Meta's Acceptable Use Policy applies, and organisations above 700 million MAU need a separate Meta licence.
How much does the Nous Portal cost?
As listed on Nous Research's portal in September 2026: a free tier at $0 covering free models, then Plus at $20/month, Super at $100/month and Ultra at $200/month, each including slightly more in credits than the price, plus pay-as-you-go top-ups roughly between $10 and $200. Tool use such as web search and image generation bills against the same credit balance.
Is Hermes better than ChatGPT or Claude?
Not on raw capability at the frontier. Leading closed models remain ahead on the hardest reasoning and broad factual recall. Hermes wins on the axes hosted models cannot compete on: you own the weights, data never leaves your network, the version never shifts underneath you, and marginal inference is free once hardware is paid for.
What hardware do I need to run Hermes?
Depends on the variant. Weights take roughly two bytes per parameter at bf16, one at 8-bit, about half at 4-bit, plus a KV cache growing with context and concurrency. A 14B or quantized 36B is realistic on a single high-VRAM GPU; a 70B needs serious single-node hardware or aggressive quantization; a 405B is multi-node.
Why is Hermes 4 built on Llama 3.1 rather than Llama 3.3?
Nous has not published a rationale, so the honest answer is structural: Meta shipped Llama 3.3 only as an instruction-tuned 70B, with no 405B counterpart. Hermes 4 was designed as one family with a shared prompt format across three sizes including a 405B, and only the Llama 3.1 generation offered a base model at that scale. Hermes 4.3 then moved off Llama entirely onto Seed-OSS-36B — which also moved it to Apache 2.0.
Is a model with reduced refusals safe to put in front of customers?
Only with guardrails you build. Lower refusal rates mean the model declines fewer legitimate professional tasks, which is often what you want internally, but nobody is moderating on your behalf. Before any customer-facing deployment you need input filtering, output moderation, prompt-injection defences, logging and access control. For internal tools used by trained staff the bar is lower, but never zero.
Can I fine-tune Hermes on my own data?
Yes, subject to the naming and attribution terms above for the Llama-based variants. In practice, try retrieval-augmented generation and prompt engineering first: fine-tuning is a real project with data preparation, training cost, evaluation and a maintenance burden every time you move to a newer base. Most problems people reach for fine-tuning to solve are retrieval problems.
How do I try Hermes without setting up infrastructure?
Three routes. Nous runs its own chat interface and the Nous Portal API. Third-party gateways list the models — as of 1 September 2026, OpenRouter carried Hermes 4 70B and 405B at 131K context. For local use, GGUF builds of Hermes 4.3 36B are published and listed on LM Studio, so it runs in llama.cpp-based desktop tools.