Hermes 3 405B vs 70B: When the Big One Is Worth It

Nous Research shipped Hermes 3 as a family, and two members of it sit next to each other in the LU Labs Cloud chat catalog: Hermes 3 70B and Hermes 3 405B. Same house style, same training approach, same absence of a thinking mode. One is several times the parameter count of the other.

Because both are in the Hosted catalog at €19 a month, the question is never "which one can I get". It is "which one do I point at this particular job", and the honest answer is that most people should be defaulting to the 70B and reaching up on purpose.

What the two share

Start with the things that do not change between them, because that is most of the list.

Both are Llama 3.1 fine-tunes. Both are instruct models with no separate reasoning pass, so neither one has a Think toggle. If you want a model that writes out its working before it answers, neither of these is the pick; Qwen 3.6 35B A3B in the same catalog has that switch.

Both call tools through the same route. Hermes was not trained on the function-calling format most APIs expect, so our proxy translates tool definitions into the prompt convention it was trained on and reads the calls back out. That means Agent mode and Code mode run on either one, and switching between them does not change how your tools behave.

Both are reachable from the same personal API key at https://lu-labs.ai/api/inference/v1 in the OpenAI chat completions shape. The only thing that changes in your code is the model id.

So the difference is not features. It is behaviour under pressure, and price.

Where the 405B actually earns its keep

The extra parameters do not make the model know more interesting facts. What they buy, in the work we have watched, is stability over length.

Give both models a system prompt with a strict output contract, say three named blocks in a fixed order with rules about what never to include. Both will follow it. Feed a long run of messages through and the smaller model starts to soften: a block goes missing, a closing line creeps back in, the "say unsure when you are unsure" rule quietly stops firing. The 405B holds the shape longer.

The same pattern shows up in persona work and in long document passes. It is not that the 70B gets things wrong and the 405B gets them right. It is that the 70B needs reminding and the 405B does not.

Three cases where that difference is worth paying for:

• A long structured pass. A large batch of records to classify against a fixed schema, where a drifted output halfway through means re-running the whole thing.
• A voice you have to hold. A character, a brand tone, a persona that has to sound the same at the start of a session and at the end of it.
• An instruction stack with exceptions. Rules that contradict each other in edge cases, where following the wrong one is worse than a clumsy sentence.

Where the 70B is the better answer

Everything else, more or less, and especially anything long-running.

Chat rates in our catalog are per output token. Hermes 3 70B costs 0.07 credits per output token, and one credit is $0.00001, so a €5 credit pack is roughly 2.3 million output tokens on it. The 405B is larger and draws more per token. That gap does not matter when you send one prompt. It matters enormously when an agent loop is generating tokens for an hour.

Agent and coding sessions are the clearest case. They produce a lot of output, most of which is intermediate scaffolding nobody reads: tool calls, file reads, plans it revises two steps later. Paying a premium per token for the model's ability to hold a persona across a long session is money spent on a quality nobody is measuring in that loop. Run the loop on the 70B.

Short questions are the other case. If the whole interaction is one prompt and one paragraph back, the smaller model's tendency to drift never gets a chance to show up, because there is nothing to drift across.

The mixed setup, which is what most people end up with

Credits are one pool. Chat, code, image and video all draw from the same wallet, and the model picker is one list. Nothing stops you from switching models inside a single project, and that is what the setup rewards.

A pattern worth copying: run the exploration and the agent loop on the 70B, then take the final artifact, the thing that ships, and re-run just that pass on the 405B with the strict system prompt. You pay the higher rate on a small fraction of the tokens and you get the stable output where it counts.

The €19 Hosted plan carries 900,000 credits a month across the whole catalog, so this kind of mixing is a budgeting decision, not an access one. And if you want to try the comparison before committing to a month, a €5 pack needs no subscription and the credits never expire. See the plans and packs for the full table.

What neither of them fixes

Neither model reasons step by step unless you ask it to in the prompt, and asking a non-reasoning model to think out loud is not the same thing as a real reasoning pass. For genuinely hard multi-step problems, pick a reasoner from the same picker.

Neither runs on your laptop. The 70B is closer to feasible than the 405B, but a 405 billion parameter model at 4-bit is hundreds of gigabytes in memory, which is a data center job and always will be. That is the honest reason Hermes 3 405B lives in the cloud rather than on your disk. Our roundup of the best open-weight models in 2026 covers what does fit at home.

And Cloud is shared hardware, so at busy hours either model can wait a few seconds in a queue before the answer starts streaming. Account data is hosted in the EU, you can delete it yourself, we never train on your data and we never sell it. Content involving minors is refused everywhere, always.

The short version

Default to Hermes 3 70B. Move to Hermes 3 405B when you have written a system prompt with real rules in it and you have watched the smaller model stop obeying them. That is the signal, and it is a signal you can actually observe rather than guess at.

The practical next step is the step by step guide to using Hermes 3 405B, which covers the picker, the API key and the prompt shape. Or start on Hosted and run both against the same system prompt for an afternoon.