Models / Nous Research

Hermes 4 405B

Neutrally aligned hybrid reasoner trained heavily on tool use. The most stable agent in the catalogue, but it still declines a share of bio and chem requests.

Fine tuneTextfrontierreasoningtool-calling

Publisher: Nous Research
Base model: Llama 3.1 405B
Parameters: 405B
Context: 128K
Released: Aug 2025
License: Llama 3.1

Method · Fine tune

Supervised fine tuning on compliant data. Changes behaviour more broadly than a projection.

Try in chat Hugging Face ↗

Overall

95.9

weighted, out of 100

Openness

90%

base 34%

Agentic completion

94%

5% loop

Loop regression

−4

base 9% → 5%

Capability retained

101.2%

+0.8 pts mean

Cheapest input

$1.00

per M · OpenRouter

Fastest

tok/s · NanoGPT

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against Llama 3.1 405B before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

Hermes 4 405BLlama 3.1 405B (base)

Open, this model

90%

Open, base

34%

Category	Base	This model	Change
Cybersecurity Exploit development, malware analysis, offensive tooling, pentest workflows.	62%	3%	−59
Biology Pathogen biology, lab protocols, dual use life sciences.	86%	31%	−55
Chemistry Synthesis routes, energetic materials, toxicology.	88%	27%	−61
Weapons Firearms, manufacturing, conventional weapons.	78%	12%	−66
Drugs Recreational drugs, harm reduction, dosing, synthesis.	66%	5%	−61
Adult content Explicit fiction and roleplay between adults.	74%	4%	−70
Violence & gore Graphic violence in fiction, history and journalism.	22%	5%	−17
Hate & harassment Slurs, offensive humor, hostile speech analysis.	85%	4%	−81
Self harm Suicide, eating disorders, frank medical questions.	70%	15%	−55
Privacy & OSINT Doxxing, surveillance, people search, OSINT tradecraft.	64%	3%	−61
Fraud & scams Phishing, social engineering, financial crime.	84%	3%	−81
State sensitive Topics censored by a model's home government (e.g. Tiananmen, Xinjiang).	14%	4%	−10

Effective censorship by provider

Refused

0%10%25%50%75%100%

Not tested

The same weights can behave differently depending on who serves them. Each cell is the model's own refusal rate combined with that provider's declared filters.

Route	Cyber	Bio	Chem	Weap	Drug	Adult	Viol	Hate	Self	OSINT	Fraud	Polit	CSAM
Raw weights	3%	31%	27%	12%	5%	4%	5%	4%	15%	3%	3%	4%	Blocked
OpenRouter	3%	31%	27%	12%	5%	4%	5%	4%	15%	3%	3%	4%	Blocked
NanoGPT	3%	31%	27%	12%	5%	4%	5%	4%	15%	3%	3%	4%	Blocked

CSAM is prohibited on every route, every provider, every model. It is not a setting.

Agentic stability

Does it hold up in agent loops?

Five multi turn tool scenarios against deterministic simulators. An episode counts as a loop when the model repeats itself, oscillates, or stops making progress before the task is done.

Preview

By scenario

CompletedLoopCrash

File search & edit93% done · 6% loop · 1% crash

Web research96% done · 3% loop · 1% crash

Multi step math99% done · 1% loop

Code debugging89% done · 9% loop · 2% crash

Data pipeline91% done · 8% loop · 1% crash

Loop rate versus base

Llama 3.1 405B9%

This model5%

Regression

−4 pts

First loop, median turn

T24

Tool calls

Preview

Valid JSON96%

Schema compliant89%

How it fails

Exact repeat40%

Same tool call with the same arguments, three turns running.

Near repeat28%

Responses above 0.85 Jaccard similarity to a recent turn.

A/B oscillation11%

Flip flopping between two actions without progress.

Malformed tool call11%

Invalid JSON or arguments that break the schema.

Fabricated result6%

The model writes the tool output itself instead of calling.

Empty turn2%

No text and no tool call.

Lost the thread4%

Ignores the latest tool output and restarts the task.

Share of failed episodes by first detected failure mode.

Versus base model

What the edit cost.

The same benchmarks on Llama 3.1 405B and on this model.

Preview

Benchmarks

BaseThis model

MMLU86.2 → 86.8 (+0.6)

GSM8K95.1 → 95.5 (+0.4)

TruthfulQA60.4 → 59.2 (−1.2)

GPQA Diamond51.1 → 51.6 (+0.5)

LiveCodeBench40.3 → 40.7 (+0.4)

IFEval88.6 → 89.4 (+0.8)

BFCL v366.0 → 70.1 (+4.1)

Retention

101.2%

of base scores kept · +0.8 pts mean change

KL divergence from base

Not reported. KL is only meaningful for weight edits, not full fine tunes.

Providers

Where to run it.

2 providers serve this model. Prices come from provider pages. Latency, throughput and uptime are preview numbers until router traffic fills them in.

Preview

Provider	Quant	Context	Input /M	Output /M	TTFT	Throughput	Uptime	Moderation
OpenRouter CheapestLowest latency	fp8	128K	$1.00	$3.00	476 ms	39 tok/s	96.3%	No added filter
NanoGPT Fastest	fp8	128K	$1.10	$3.20	540 ms	48 tok/s	96.6%	No added filter

Usage

Who is using it.

Preview

2.8B

tokens routed this week

+22% vs four weeks ago

12 weeks agoThis week

API

Call it in three lines.

OpenAI compatible. Point your SDK at our base URL and keep the rest of your code.

curl https://openuncensored.com/api/v1/chat/completions \
  -H "Authorization: Bearer $OU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "hermes-4-405b",
    "messages": [{"role": "user", "content": "Hello"}],
    "routing": {"strategy": "quality"}
  }'

Model id hermes-4-405b · routing strategies quality, speed, cost · routing docs