Open-weight 70B hybrid reasoner you can toggle between direct answers and explicit <think> traces, with low refusal rates.
Modality:
Text
PDF
model ID
nousresearch/hermes-4-70b
Output Speed
74.07
tok/s
Intelligence Index
9.9
/ 100
Context Window
131072
tokens
Input price
0.78
Anytoken
Output price
2.4
Anytoken
Hermes 4 70B: A Steerable, Toggleable-Reasoning Open-Weight Model
Hermes 4 70B is an open-weight hybrid reasoning model from Nous Research, built on Meta-Llama-3.1-70B. It sits below the 405B flagship and above the 14B variant in the Hermes 4 lineup. Its defining trait is a hybrid mode: the model can answer directly or emit explicit … reasoning traces, toggled per request. Trained on an expanded ~60B-token post-training corpus emphasizing verified reasoning, it targets math, code, STEM, logic, and structured outputs while maintaining low refusal rates and strong instruction adherence. It suits teams wanting steerable, neutrally-aligned generation with optional test-time compute.
Get an API key and start calling Hermes 4 70B
Performance
Toggleable Test-Time Compute for Math, Code and Logic
Hermes 4 70B is strongest on quantitative reasoning when its thinking mode is enabled. In reasoning mode it reaches 95.6 on MATH-500 and 73.5 on AIME'24 in Nous Research's own evaluation. Because reasoning is a per-request toggle rather than an always-on cost, teams can reserve expensive deliberation for hard math, code, or logic prompts and answer routine requests directly. In production this means one checkpoint serving both fast and deliberate paths, controlling latency and token spend based on task difficulty rather than provisioning two separate models.
Benchmarks
Where Hermes 4 70B's Numbers Actually Come From
The headline scores are self-reported in Nous Research's technical report, measured with reasoning enabled. Hermes 4 70B posts 95.6 on MATH-500 and 73.5 on AIME'24. The family's signature result is RefusalBench, an internal Nous benchmark of 166 hand-crafted prompts judged by Sonnet 4, where Hermes 4 scored 57.1 in reasoning mode — far above the ~17% reported for GPT-4o and Claude Sonnet 4 on the same test. Note these are provider evaluations run in reasoning mode; hosted endpoints default to non-reasoning, so real-world numbers depend heavily on whether the <think> toggle is enabled.
Output Speed
74.07
tok/s
Intelligence Index
9.9
/ 100
MMLU
Broad world knowledge and problem-solving
81
%
GPQA
PhD-level scientific reasoning across physics, biology, chemistry.
70
%
HLE
Adherence to multi-step structured instructions.
9
%
LiveCodeBench
Tool-calling reliability in long agentic loops.
65
%
Technical Specifications
What the model supports
Hermes 4 70B is a text-only model with a 131,072-token context window shared across the Hermes 4 family. The most consequential specification is the hybrid reasoning toggle: reasoning can be switched on or off per request, so the model does not force chain-of-thought overhead onto simple prompts. It is trained for schema adherence, valid JSON, function calling and tool use, though tool availability depends on the specific hosting endpoint. As an open-weight model on a Llama-3.1 base, it supports self-hosting and fine-tuning subject to the underlying license.
Verified Specifications —
Hermes 4
Input modalities
Text
PDF
output modalities
Text
Context window
131072
tokens
Maximum output tokens
0
Reasoning
Yes
Knowledge cutoff
August 2025
Pricing (standard)
0.78
AnyTokens in
/
2.4
AnyTokens out
Quickstart
import requests
url = "https://api.anyapi.ai/v1/chat/completions"
payload = {
"stream": False,
"tool_choice": "auto",
"logprobs": False,
"model": "Model_Name",
"messages": []
}
headers = {
"Authorization": "Bearer AnyAPI_API_KEY",
"Content-Type": "application/json"
}
response = requests.post(url, json=payload, headers=headers)
print(response.json())import requests url = "https://api.anyapi.ai/v1/chat/completions" payload = { "stream": False, "tool_choice": "auto", "logprobs": False, "model": "Model_Name", "messages": []
}
headers = { "Authorization": "Bearer AnyAPI_API_KEY", "Content-Type": "application/json"
} response = requests.post(url, json=payload, headers=headers) print(response.json())
View docs
Copy
Code is copied
const url = 'https://api.anyapi.ai/v1/chat/completions';
const options = {
method: 'POST',
headers: {Authorization: 'Bearer AnyAPI_API_KEY', 'Content-Type': 'application/json'},
body: '{"stream":false,"tool_choice":"auto","logprobs":false,"model":"Model_Name","messages":[]}'
};
try {
const response = await fetch(url, options);
const data = await response.json();
console.log(data);
} catch (error) {
console.error(error);
}const url = 'https://api.anyapi.ai/v1/chat/completions';
const options = { method: 'POST', headers: {Authorization: 'Bearer AnyAPI_API_KEY', 'Content-Type': 'application/json'}, body: '{"stream":false,"tool_choice":"auto","logprobs":false,"model":"Model_Name","messages":[]}'
}; try { const response = await fetch(url, options); const data = await response.json(); console.log(data);
} catch (error) { console.error(error);
}
View docs
Copy
Code is copied
curl --request POST \
--url https://api.anyapi.ai/v1/chat/completions \
--header 'Authorization: Bearer AnyAPI_API_KEY' \
--header 'Content-Type: application/json' \
--data '{
"stream": false,
"tool_choice": "auto",
"logprobs": false,
"model": "Model_Name",
"messages": []
}'curl --request POST \ --url https://api.anyapi.ai/v1/chat/completions \ --header 'Authorization: Bearer AnyAPI_API_KEY' \ --header 'Content-Type: application/json' \ --data '{ "stream": false, "tool_choice": "auto", "logprobs": false, "model": "Model_Name", "messages": [] }'
View docs
Copy
Code is copied
View docs
Comparison
Hermes 4 70B vs Hermes 4 405B: Which Size Do You Deploy?
Both models share the same architecture family, the same 131K context window, the same hybrid reasoning toggle, and the same neutrally-aligned, low-refusal training philosophy. The 405B is built on Llama-3.1-405B and the 70B on Llama-3.1-70B. The practical decision is compute versus quality: the 405B reaches higher reasoning ceilings (96.3 MATH-500, 81.9 AIME'24 reported) while the 70B trails only modestly (95.6 MATH-500, 73.5 AIME'24) at a fraction of the memory footprint and cost, making it far cheaper to run at volume.
Dimension
Hermes 4
Hermes 4
Context window
131072
tokens
131072
tokens
Output speed
74.07
tok/s
42.74
tok/s
Intelligence Index
9.9
8.8
Input pricing
0.78
AnyToken
6
AnyToken
Output pricing
2.4
AnyToken
18
AnyToken
Knowledge cutoff
August 2025
August 2025
Choose Hermes 4 70B when you need the same steerability and hybrid reasoning behavior at lower cost and lighter hardware, for high-volume or latency-sensitive workloads where the modest accuracy gap is acceptable. Choose Hermes 4 405B when you need the maximum reasoning ceiling for the hardest math, code, or logic tasks and can absorb the higher per-token cost and larger serving footprint. Many teams route easy traffic to 70B and escalate difficult prompts to 405B.
Limitations & Trade-offs
1
Self-reported benchmarks in reasoning mode. The headline MATH-500 and AIME scores come from Nous Research's own technical report with reasoning enabled. Many hosted endpoints default to non-reasoning mode, and one independent leaderboard ranked the 405B mid-pack precisely because it was benchmarked without reasoning. If you deploy Hermes 4 70B without enabling the <think> toggle, expect materially lower quality than the published numbers, so validate on your own tasks in the exact mode you intend to run.
2
Text-only, no multimodality. Hermes 4 70B accepts and produces text only. It has no image, audio, video, or document-vision input. Workloads requiring screenshot understanding, chart reading, PDF layout parsing, or any visual grounding need a different model. This limits it for document-processing pipelines that depend on native vision rather than pre-extracted text.
3
Endpoint-dependent tool calling. The model is trained for function calling, tool use, and schema-adherent JSON, but these capabilities are not uniformly exposed. Some hosted endpoints do not accept tools, disabling function calling entirely, and offer JSON response formatting without strict schema enforcement. For agentic workflows that rely on guaranteed tool invocation or hard schema validation, confirm the specific endpoint's capabilities before committing.
4
Neutral alignment shifts safety responsibility to you. Hermes 4's low refusal rate (57.1 on RefusalBench versus ~17% for GPT-4o and Claude) is intentional: the model is steered by the user and system prompt rather than a built-in ethics policy. That makes it flexible for research and unrestricted creative work, but it means content filtering, guardrails, and policy enforcement must be implemented in your own application layer for customer-facing products.
Best-Fit Workloads
Where this model earns its place
01
Math and STEM problem solving
With reasoning enabled, Hermes 4 70B is a strong fit for math-heavy and STEM tutoring, verification, and step-by-step derivation tasks, backed by 95.6 on MATH-500 and 73.5 on AIME'24 in the provider's evaluation. The hybrid toggle lets you spend tokens on deliberation only for genuinely hard problems, keeping routine queries cheap. Validate accuracy in your target domain since scores reflect reasoning mode.
02
Structured data and JSON generation
The model is trained specifically for schema adherence and can produce and repair valid JSON. This suits extraction pipelines, config generation, and tool-argument formatting where format faithfulness matters more than raw creativity. Confirm whether your endpoint enforces strict JSON schemas, as some expose only response_format without schema validation.
03
Steerable creative writing and roleplay
Hermes 4's neutral alignment and low refusal rate make it well suited to creative writing, character roleplay, and open-ended generation where boilerplate lecturing and sycophancy are unwanted. Nous explicitly positions the model for a more humanistic, user-directed interaction. This flexibility requires you to add your own content controls for public-facing deployments.
04
Cost-sensitive high-volume text automation
As the mid-size open-weight variant, Hermes 4 70B occupies a lower-cost tier than the 405B while retaining the same hybrid reasoning and steerability. That makes it a practical default for high-volume internal automation, summarization, and drafting where reasoning can stay off for most requests and be selectively enabled. Its 131K context supports long documents in a single pass.
Integration
01
One API integration
02
Easy model switching
03
Flexible for production
04