Hermes 4 vs Hermes 3: What Changed and Which to Pick

What Changed Between Hermes 3 and Hermes 4

Hermes 3, released by Nous Research in 2024, was a family of full fine-tunes of Llama 3.1 at 8B, 70B, and 405B. Its focus was neutral alignment and steerability: a strong, uncensored assistant that would do what you asked without adding opinions of its own. It did not reason step by step by default. It answered.

Hermes 4, released in August 2025, is a different animal. It ships in three sizes (14B on Qwen 3, 70B and 405B on Llama 3.1) and adds a hybrid reasoning mode where the model chooses whether to think in <think>...</think> traces before answering. The post-training corpus grew roughly fifty times: from about 1M samples and 1.2B tokens for Hermes 3 to about 5M samples and 60B tokens for Hermes 4, blended across reasoning and non-reasoning data. Training ran on 192 NVIDIA B200 GPUs.

Then, in December 2025, Nous released Hermes 4.3 36B. It is based on ByteDance's Seed-OSS-36B architecture, extends context to 512K tokens, and was post-trained end to end on Nous's own decentralized Psyche network. It nearly matches Hermes 4 70B on evaluations at half the parameter count.

If you run Hermes Agent with your own OpenRouter or direct-Nous key, the model choice is now genuinely interesting. Here is what actually differs and how to pick.

The Hermes 4 Lineup at a Glance

Model	Base	Reasoning Mode	Best For
Hermes 4 14B	Qwen 3 14B	Yes	Local inference, edge devices, low-cost drafts
Hermes 4 70B	Llama 3.1 70B	Yes	Cost-conscious BYOK workloads that need real quality
Hermes 4 405B	Llama 3.1 405B	Yes	Frontier reasoning: math, code, hard STEM questions
Hermes 4.3 36B	Seed-OSS 36B	Yes	512K context work, local self-hosting on one GPU

All four share the same hybrid reasoning trick: one set of weights, one toggle. You do not maintain a separate "reasoning" model and a separate "instruct" model.

Hybrid Reasoning: The Feature That Matters Most

In Hermes 3 you got a straight response. Fast, competent, done.

In Hermes 4 you can prompt for either behavior on the same checkpoint. When reasoning mode is active, the model emits a <think>...</think> block containing its internal deliberation, then a final answer. When it is off, the model answers directly, like Hermes 3 did.

Two implications for how you use it inside Hermes Agent:

You do not need two provider slots. Point BYOK at one Hermes 4 model and switch modes per task by system prompt or wrapper convention. Cheaper and simpler than routing between separate reasoning and non-reasoning models.
You pay for the tokens the model spent thinking. Reasoning mode is not free. A long <think> block can easily double the output cost of a hard question, so keep it off for routine summarization or chit-chat.

If you want a deeper feel for how a reasoning agent behaves in day-to-day use, our post on Hermes Agent memory and skills walks through how memory files change what the agent chooses to reason about.

Benchmarks: What the Numbers Actually Show

At the top of the stack, Hermes 4 405B in reasoning mode reports:

MATH-500: 96.3
AIME'24: 81.9
GPQA Diamond: 70.5
LiveCodeBench: 61.3

Those are competitive with DeepSeek R1 671B and Qwen 3 235B, on open weights. The MATH-500 headline is real, but treat it as a ceiling for a particular kind of question rather than a promise about every prompt: on non-reasoning tasks the same model behaves closer to a strong tuned Llama 3.1 405B than to a frontier reasoner.

Hermes 4 70B trades several points on the hard reasoning benchmarks for roughly six times cheaper inference. Hermes 4 14B is the sleeper: it is small enough to run on a single consumer GPU while keeping the hybrid reasoning behavior, which is unusual at that size.

Hermes 4.3 36B is the most interesting one to watch. Nous reports it nearly matches Hermes 4 70B at half the parameters, thanks to the larger Seed-OSS base and the distributed Psyche post-training run. It also has a 512K context window, which the 70B does not.

Which One Should You Run Through Hermes Agent

Hermes Agent is model-agnostic. You choose in your config.yaml or through the Hermes Agent OpenRouter setup, and the harness handles the rest. Here is a rule of thumb for BYOK users:

Long-context summarization, long chat memory, code review of large files. Pick Hermes 4.3 36B. The 512K window is the deciding factor. Nothing else in the family competes there.
Hard reasoning: math, novel code, agentic planning with many steps. Pick Hermes 4 405B in reasoning mode. Turn reasoning off for the routine turns and back on for the hard ones.
Cost-conscious daily driver. Pick Hermes 4 70B. It is the balanced default. Good answers, reasonable price, still ships hybrid reasoning if you need it.
Edge, local, or offline runs. Pick Hermes 4 14B, or Hermes 4.3 36B if you have a beefier GPU. The 14B GGUFs run comfortably on a single mid-range card.
You have a working Hermes 3 setup and it fits your workload. Do not migrate for its own sake. Hermes 3 405B is still a strong, non-reasoning assistant. Move only when you need one of the four capabilities above.

For a broader take on picking a provider layer, see our comparison of Hermes Agent hosting vs self-hosting. The model choice is largely orthogonal to where you run the agent.

How to Swap Models in Hermes Agent

The switch is a config change, not a rebuild. In ~/.hermes/config.yaml:

model:
  provider: openrouter
  name: nousresearch/hermes-4-405b
  # For direct Nous access, use:
  # provider: nous
  # name: hermes-4-405b

Restart the gateway:

hermes gateway restart

For reasoning mode, most Hermes 4 models accept a reasoning: true or enable_thinking: true flag on the provider side, or you can wrap your prompt with an explicit <think> directive at the system-prompt level. Check your provider's docs before assuming defaults.

If you run BYOK through a managed host, the switch is even simpler: change the model string in your dashboard and send a new message. No restart, no image rebuild.

The Bottom Line

Hermes 3 to Hermes 4 is a bigger jump than the version number suggests. You go from a strong Llama-3.1 fine-tune with steerability to a family of models with hybrid reasoning, competitive frontier benchmarks, a distributed training pipeline (4.3), and a genuine long-context option. The migration is a config change if you already run Hermes Agent, and the ROI is highest for workloads that involve real reasoning, long context, or a mix of both.

Pick the smallest model in the family that answers your hardest question well. Turn reasoning on only when you need it. And keep an eye on the Psyche network releases: the 4.3 pattern (smaller base, larger post-training, decentralized run) is almost certainly the shape of the next few Hermes releases.

Get started with Hermify if you want the model swap to be a dropdown instead of a config file, and a Hermes Agent running on Telegram while you decide.