Hermes-4-405B: specs, benchmarks, and how to run it on Lyceum

Get started: call Hermes-4-405B on Lyceum

You can access Hermes-4-405B through Lyceum Technology's OpenAI-compatible Serverless Inference API. Because the API is a drop-in replacement, you only need to update your base URL and API key to start querying the model on European infrastructure.

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-4-405B",
 messages=[{"role": "user", "content": "Hello!"}],
 max_tokens=256,
 )
 print(response.choices[0].message.content)

Pricing and region for Hermes-4-405B

This model is hosted in the eu-north1 region, so requests to it are processed in European data centres, with no training on customer data and no data retained after processing. Serverless Inference has no service tiers: the model you pick sets the balance between latency and capability, and Hermes-4-405B is built for high-capability workloads requiring maximum precision.

• Input pricing: $1.00 per million tokens
• Output pricing: $3.00 per million tokens

By utilizing our per-token billing model, you avoid the massive overhead of provisioning dedicated 8x H100 clusters for a 405B parameter model. You only pay for the exact compute you consume, scaling instantly from zero to peak traffic without managing underlying hardware. The snippet above demonstrates a standard chat completion request. For advanced use cases, you can also pass custom system prompts to trigger the model's hybrid reasoning mode, instructing it to deliberate before returning the final response.

What Hermes-4-405B is good at

Hybrid reasoning and deep deliberation

Hermes-4-405B introduces a hybrid reasoning mode. When faced with complex problems, the model can output <think>...</think> tags to deliberate internally before providing a final answer. This chain-of-thought processing improves performance on advanced mathematics, logic puzzles, and multi-step coding tasks. Because it is a hybrid model, developers retain control: you can prompt it to think deeply for complex tasks, or bypass the reasoning tokens entirely for fast, direct responses when latency is the priority.

Steerability and reduced refusals

Unlike proprietary models that often refuse benign requests due to overly strict safety tuning, Hermes-4-405B is designed to be highly steerable and aligned to the user. It scores 57.1 on Nous Research's internal RefusalBench, the highest of the open-weight models compared in the Hermes 4 technical report, meaning it follows instructions faithfully without injecting unwanted moralistic biases or unnecessary censorship. This makes it ideal for creative writing, nuanced roleplay, and enterprise applications where unpredictable refusals break the user experience.

Structured outputs and schema adherence

The model was explicitly trained on a large post-training corpus to produce valid JSON and adhere to strict schemas. It can even repair malformed JSON objects autonomously. This makes Hermes-4-405B a reliable engine for agentic workflows, function calling, and data extraction pipelines where format fidelity is critical. By combining deep reasoning with strict formatting, developers can build robust AI agents that reliably interact with external APIs and databases without requiring complex retry logic.

Limitations and trade-offs

Where Hermes-4-405B falls short

While Hermes-4-405B is highly capable, its 405-billion parameter size introduces specific trade-offs for production deployments.

First, the hybrid reasoning mode, while powerful, consumes significantly more output tokens. Because the model generates internal monologue inside <think> tags before answering, the time-to-first-token (TTFT) for the actual response is delayed, and the total cost per request increases due to the higher output token volume. For simple classification or basic summarization tasks, this deep deliberation is overkill and less cost-effective than using a smaller, faster model like Hermes-4-70B.

Second, as a dense 405B model, it requires substantial VRAM to run. If you choose to host it on dedicated infrastructure rather than using the serverless API, you will need an 8x H100 or 8x B200 cluster just to load the weights and maintain a usable KV cache. This makes dedicated deployments expensive for teams with low or bursty traffic, making the serverless per-token option much more viable.

Finally, while its lack of censorship is a feature for many developers, enterprise teams building public-facing chatbots may need to implement their own guardrails. Because the model is highly steerable and will not automatically refuse controversial prompts, you must rely on system prompts or external moderation layers to ensure brand safety in customer-facing applications.

Benchmarks and how it compares

Hermes-4-405B benchmark results

Hermes-4-405B competes directly with frontier proprietary models, particularly when its reasoning mode is active. It shows improvements over the Llama-3.1 base model in STEM and coding evaluations, achieving high scores on advanced mathematical datasets.

All three figures are from Table 3 of the Hermes 4 Technical Report (arXiv:2508.18255), reasoning mode, read 3 August 2026.

Comparison to sibling models

Compared to Hermes-4-70B, the 405B variant scores higher on graduate-level science questions (GPQA Diamond, 70.6 against 66.1) and on coding (LiveCodeBench v6, 61.4 against 50.5), while the two are level on MMLU-Pro at 80.6 and 80.7. However, the 70B model is much faster and cheaper to run, making it the better choice for high-volume, low-latency applications where frontier-level reasoning is not strictly required. The smaller model in the same family is covered in our notes on Hermes-4-70B and how to run it.

When compared to the base Llama-3.1-405B-Instruct, Hermes-4-405B provides much stronger schema adherence and a significantly lower refusal rate. The base Llama model often struggles with overly cautious safety alignments that trigger false refusals on benign coding or creative tasks. Hermes-4-405B strips away these limitations, making it far more suitable for agentic workflows and automated tool use where predictable execution is mandatory. For developers building autonomous systems, this reliability translates directly to fewer failed API calls and less complex retry logic.

Using it in production

Production configuration for Hermes-4-405B

When deploying Hermes-4-405B via Lyceum Technology's Serverless Inference API, understanding the context window and pricing structure is essential for optimizing your workloads.

The model supports a massive 128k token context window, allowing you to pass entire code repositories, long financial documents, or extensive multi-turn conversation histories in a single prompt. Because Serverless Inference has no tiers, the latency profile follows from the model itself: Hermes-4-405B is served in the eu-north1 region and is built for high-capability tasks requiring maximum precision rather than pure speed.

To calculate costs, consider a typical agentic workflow: analyzing a 10,000-token document and generating a 2,000-token structured JSON report (including reasoning tokens). At $1.00 per million input tokens and $3.00 per million output tokens, this request would cost $0.01 for the input and $0.006 for the output, totaling $0.016 per execution.

If you are using the hybrid reasoning capabilities, we strongly recommend enabling streaming (stream=True in the OpenAI SDK). Because the model generates extensive <think> tokens before producing the final answer, streaming prevents your application from timing out and provides immediate feedback to the user while the model deliberates. You can parse the stream on the client side to hide the <think> tags from the end user while still benefiting from the model's enhanced logical reasoning.

Running Hermes-4-405B on European infrastructure

Why run Hermes-4-405B on Lyceum

For European enterprises and AI startups, data privacy is often the primary blocker to adopting frontier models. While US-based providers offer similar APIs, they route data outside the EU, creating significant GDPR and compliance risks. Non-EU hosting is frequently a deal-breaker for teams handling sensitive healthcare, financial, or proprietary manufacturing data.

When you query the model via our API, requests to Hermes-4-405B are processed in the eu-north1 region, in European data centres, with no training on customer data and no data retained after processing. Lyceum's position is GDPR-compliant processing on a self-asserted basis: Lyceum holds no ISO 27001 or SOC 2 certificate today and makes no EU AI Act conformity claim, and the data centre operators hold certifications at facility level.

Furthermore, our open-stack transparency means you avoid vendor lock-in. Because we use an OpenAI-compatible API, migrating your existing applications to Lyceum takes minutes. You get the reasoning power of a 405-billion parameter model, the cost-efficiency of per-token billing, and processing in European data centres, all without managing a single GPU.

By leveraging our serverless inference platform, you eliminate the need to provision expensive 8x H100 clusters. You pay for the tokens you use, scaling from zero to peak demand while your requests are processed in European data centres.

Sources

[1] NousResearch/Hermes-4-405B - Hugging Face; [2] Hermes 4 Technical Report - arXiv; [3] unsloth/Hermes-4-405B-GGUF - Hugging Face