Hermes 2.3: what's new in Nous Research's model
Hermes 2.3 by Nous Research: the new features, the benchmarks, the use cases, and where to run it cheapest.
~ 12 min read
Hermes is the line of open models fine-tuned by Nous Research, a lab known for producing some of the best open-weights instruct models. The 2.3 generation landed recently, and it's worth understanding what's new — especially if you compare inference providers on this site, because Hermes is one of the most-run open models.
What is Hermes, in one paragraph
Nous Research takes strong base models (Llama, Mistral, Qwen) and fine-tunes them heavily on curated datasets, with an emphasis on instruction following, character/persona, and open-ended reasoning — the opposite of the "polished but bland" style of some big labs. The result is a model that's often considered more "creative" and better at nuanced instructions than the base it came from. The Hermes name has become a quality signal in the open community.
What's new in 2.3
- Refined instruction following: better on the IFEval benchmark (precise instruction adherence) — the area where the 2.x line already shone.
- Reduced hallucination: Nous worked on factual grounding, notably for agentic use where the model must not invent tool outputs.
- Improved tool use & function calling: more reliable structured output (JSON), which matters for agents and pipelines.
- New sizes: the 2.3 release refreshed the lineup on recent bases (Qwen 2.5, Llama 3.x), keeping the 8B / 70B / 405B tiers.
- Long context: support for long contexts carried over from the base, with better coherence over long conversations.
Indicative benchmarks (Nous Research releases)
| Benchmark | Hermes 2 Pro | Hermes 2.3 | Delta |
|---|---|---|---|
| IFEval | ~83 | ~86 | +3 |
| MMLU | ~79 | ~80 | +1 |
| HumanEval | ~73 | ~75 | +2 |
| GSM8K | ~88 | ~89 | +1 |
| TruthfulQA | ~59 | ~62 | +3 |
Read this as "consistent incremental gains," not a generational leap. The big news is the reliability of structured output and the reduced hallucination for agentic use — not a raw-knowledge jump.
Where it shines
- Agents & tool use: reliable function calling and JSON — a major plus for pipelines.
- Character & creative writing: the Nous house style — more "voice" than the sterile output of some frontier models.
- Open-ended reasoning: less "refusal" and more willingness to engage.
- Open weights: runnable locally — check the vocabulary guide for sizes and the hardware guide for what fits.
Where to run it, and at what price
Hermes is available on most aggregators (OpenRouter, Together, DeepInfra, Fireworks) and on the Nous Research Hugging Face org for local download. The price comparator on the home page is the right tool to find the cheapest provider for the size you want.
Indicative prices (per million tokens, USD):
| Variant | Cheapest API (approx.) | Local VRAM (Q4) |
|---|---|---|
| Hermes-3-Llama-3-8B | ~$0.10 / M | ~5 GB |
| Hermes-3-Llama-3-70B | ~$0.60 / M | ~40 GB |
| Hermes-3-Llama-3-405B | ~$3 / M | ~240 GB (Mac 192 GB tight / multi-GPU) |
For whom?
- Builders of agents: the structured-output reliability is a real plus.
- Creative / RP use: the Nous style is a known reference there.
- Local-AI tinkerers: the 8B runs on a laptop, the 70B on a Mac mini M4 Pro 64 GB or a 4090 rig.
- Those who want maximum raw knowledge: a frontier model (GPT-5, Claude) still wins — Hermes trades some breadth for character and instruction adherence.
In summary
- Hermes 2.3 = an incremental but solid refresh, focused on reliability for agents and reduced hallucination.
- Open weights → runnable locally and on aggregators at competitive prices.
- Not a raw-knowledge leap; a style & instruction leap.
To find the cheapest provider for the exact size you want, use the price comparator. To understand how models like Hermes are evaluated, read the evaluation harness guide.