Hermes-4-70B: specs, benchmarks, and how to run it on Lyceum

Get started: call Hermes-4-70B on Lyceum

To integrate Hermes-4-70B into your application, use the standard OpenAI Python SDK. Because Lyceum provides an OpenAI-compatible API, you only need to update the base URL and provide your Lyceum API key. The model string for this endpoint is NousResearch/Hermes-4-70B.

from openai import OpenAI
 
 client = OpenAI(
 base_url="<the base URL shown in your Lyceum dashboard>",
 api_key="<your lyceum api key>",
 )
 response = client.chat.completions.create(
 model="NousResearch/Hermes-4-70B",
 messages=[{"role": "user", "content": "Hello!"}],
 max_tokens=256,
 )
 print(response.choices[0].message.content)

Pricing and region for Hermes-4-70B

Lyceum serves Hermes-4-70B through Serverless Inference, which bills per token with no base fee. The pricing is $0.13 per million input tokens and $0.40 per million output tokens. Billing follows the tokens your application actually sends and receives, and there is no charge when your application is idle.

All API requests for this model are processed in the eu-north1 region. Requests to this model are processed in European data centres. This setup provides a clear path to GDPR compliance for enterprise applications, healthcare platforms, and financial services that cannot route sensitive user data through US-based API providers. The combination of these economics and European hosting makes this model practical for production deployments.

What Hermes-4-70B is good at

Hybrid reasoning and structured outputs

Hermes-4-70B introduces a hybrid reasoning mode that allows the model to deliberate before generating a final response. When faced with complex logic, the model can output explicit thinking segments to work through the problem step by step. For simpler queries, it can bypass this deliberation to provide faster responses. Furthermore, Nous Research trained the model to produce valid JSON for given schemas, making it reliable for programmatic function calling.

Steerability and reduced refusals

One of the primary design goals of the Hermes series is user alignment without excessive censorship. Hermes-4-70B scores 59.5 on RefusalBench in reasoning mode, against 15.3 for Cogito 70B in the same table, demonstrating a willingness to be helpful across scenarios that other models often block. This steerability means developers can rely on the model to follow system prompts accurately and maintain complex roleplay instructions without triggering false-positive safety refusals.

Math, code, and logic capabilities

The model was post-trained on a synthesized corpus of approximately 5 million samples, around 60 billion tokens, blended across reasoning and non-reasoning data. This dataset yields improvements in STEM fields. Hermes-4-70B excels at competitive programming tasks, advanced mathematical problem solving, and scientific reasoning. It retains the general assistant quality of its base architecture while pushing the boundaries of what a 70-billion parameter model can achieve in specialized domains.

Limitations and trade-offs

Where Hermes-4-70B falls short

While Hermes-4-70B is capable, it is constrained by its 70-billion parameter architecture. For complex reasoning or massive multi-document synthesis, larger models like the 405B variant of Hermes 4 will yield better accuracy. The 70B model strikes a balance between performance and inference cost, but it cannot match the raw parameter depth of frontier-class models.

Another trade-off involves the hybrid reasoning mode. If the reasoning mode is permanently enabled without careful prompt engineering, the model may over-deliberate on simple queries. This unnecessary thinking process consumes additional output tokens, which increases latency and drives up inference costs. Developers must actively manage when to trigger reasoning traces to maintain optimal application speed.

While the model is steerable and features reduced refusal rates, this lack of strict censorship requires developers to implement their own safety guardrails. Applications deployed in highly regulated environments will need secondary moderation layers to ensure the model does not generate inappropriate content. Teams migrating from heavily filtered models will need to adjust their system prompts to account for Hermes-4-70B's willingness to comply with almost any user instruction. Additionally, while the 131,072-token context window is substantial, performance can degrade when the context is fully saturated. Extreme edge cases require chunking strategies.

Benchmarks and how it compares

Hermes-4-70B benchmark results

Nous Research published benchmark data for Hermes-4-70B in the Hermes 4 technical report, covering mathematics, logic, coding, and general knowledge tasks. The figures below are the reasoning-mode results, where the model competes with or outperforms other open-weight models in its weight class.

Source: Hermes 4 Technical Report, Table 4, reasoning-mode scores, read 3 August 2026.

When compared to its base model, Llama-3.1-70B, Hermes-4-70B shows improvements in structured output generation and mathematical reasoning. The addition of the hybrid reasoning mode allows it to score higher on complex logic evaluations like MATH-500 and AIME 2025. Against current sibling models in the Lyceum catalogue, such as standard instruction-tuned 70B models, Hermes-4-70B offers a distinct advantage for developers who need strict JSON schema adherence and the ability to toggle deep thinking traces. Its LiveCodeBench v6 score of 50.5% in reasoning mode makes it a solid choice for coding assistants. The larger model in the same family is covered in our notes on Hermes-4-405B and how to run it.

Using it in production

Production configuration for Hermes-4-70B

When deploying Hermes-4-70B in production, understanding the model parameters and pricing structure is critical. The model supports a context window of 131,072 tokens, which is ideal for analyzing large codebases or maintaining long multi-turn conversations.

Serverless Inference has no service tiers: latency and capability follow from the model you choose, and Hermes-4-70B is built for high-throughput workloads where latency and unit economics are the primary concerns. The pricing is set at $0.13 per million input tokens and $0.40 per million output tokens.

To understand the production economics, consider an application processing 10,000 requests per day. If an average request contains 1,500 input tokens and generates 500 output tokens, the daily token volume would be 15 million input tokens and 5 million output tokens.

• Input cost: 15 million tokens × $0.13 = $1.95
• Output cost: 5 million tokens × $0.40 = $2.00
• Total daily cost: $3.95

This per-token pricing model ensures you pay for the exact compute used. Furthermore, Lyceum does not charge any egress fees, meaning you can stream large volumes of generated text back to your application without incurring hidden network transfer costs. All API requests for Hermes-4-70B are routed through the eu-north1 region, ensuring low latency for European users while maintaining strict data residency.

Running Hermes-4-70B on EU-sovereign infrastructure

Why run Hermes-4-70B on Lyceum

For European engineering teams, data sovereignty is a hard requirement. Running Hermes-4-70B on Lyceum ensures that your inference workloads remain entirely within the European Union. Because the model is hosted in the eu-north1 region, your data never crosses the Atlantic, providing a clear path to GDPR-compliant inference. This is a critical advantage over US-based API providers that route traffic through American data centers.

Lyceum serves inference from European data centres in Spain, Paris and the Nordics. By utilizing our open-stack transparency, powered by vLLM and NVIDIA Dynamo, developers gain deep visibility into the inference process. You are not locked into a proprietary engine.

The platform supports engineering velocity. The OpenAI-compatible API means you can migrate existing applications to Lyceum by changing a single line of code, the base URL. There is no need to rewrite your application logic or learn a new SDK. Additionally, you get per-token billing with no base fee, and no charge when your application is idle. Lyceum provides the performance, compliance, and cost-efficiency required to scale AI applications across Europe.

Sources

[1] NousResearch/Hermes-4-70B - Hugging Face; [2] Releases - NOUS RESEARCH; [3] Hermes 4 Technical Report - arXiv