Hermes 4 405B

This Dev.to analysis article from April 2026 critically examines the benchmark framing of Hermes 4 405B, the flagship open-weights model from Nous Research built on the Llama-3.1-405B architecture and released on August 26, 2025. The article highlights the model's hybrid reasoning mode, which lets a single checkpoint s The piece also surfaces quantitative evidence relevant to production use: Hermes 4 405B scores 81.9% on AIME 2024 and 78.1% on AIME 2025 when reasoning is enabled, and 57.1% on RefusalBench, reflecting a deliberate 'neutral alignment' stance with reduced refusal rates. Artificial Analysis ranked the model 18 on its Int

The GTM Directory describes Hermes 4 405B as a 405-billion-parameter open-weights chat and reasoning model from Nous Research, fine-tuned from Meta-Llama-3.1-405B with a 131,072-token context and an August 31, 2024 knowledge cutoff. Its standout capability is the hybrid reasoning mode that lets the user or the model toggle between fast direct answers and extended chain-of-thought, backed by a roughly 60-billion-token post-training corpus emphasizing math, code, and STEM. For builders, the page highlights wiring Hermes 4 405B into function-calling and structured-output pipelines where strict JSON schema adherence and low refusal rates are needed, and into RAG workflows where chain-of-thought traces add value on hard STEM or code tasks. Tradeoffs cited are the heavy GPU budget implied by 405B parameters, narrow multi-provider API redundancy via a single OpenRouter route at roughly 41 tokens per second, and the August 2024 knowledge ceiling.

The Corx-labs benchmark directory for Hermes 4 405B confirms the model as a Nous Research open-weights dense transformer released in August 2025, with about 405B parameters, a 131,072-token context window, up to 32,768 output tokens, text input and output, reasoning, tool calling, and the Llama 3.1 Community License. A representative third-party hosting price is listed at $1.00 per million input tokens and $3.00 per million output tokens. The same directory shows published reasoning and math scores, specifically GPQA Diamond at 70.5% (rank 35 of 83 reporting models) and AIME 2025 at 78.1% (rank 36 of 56 reporting models), with most other benchmarks marked as not reported. The summary highlights the model's hybrid reasoning toggle and reduced refusal behavior as the primary differentiators in tool-use and schema-constrained workflows.

An Artificial Analysis evaluation of Hermes 4 Llama-3.1 405B (Reasoning) scores it 7 on the Intelligence Index, placing it at rank 102 of 116, while output throughput is measured at about 40.9 tokens per second and total parameters at 406B under the Meta Llama 3 Community License. Reported pricing is $1.00 per million input tokens and $3.00 per million output tokens, with the page explicitly flagged as a third-party hosting rate rather than a Nous-published list price. The model's context window is listed at 128k tokens, which is slightly below the 131,072 figure cited elsewhere. The dashboard also confirms a reasoning-capable variant released in August 2025 with text-in/text-out support and open weights hosted on Hugging Face, and notes a non-reasoning sibling may exist. Despite sub-par intelligence and speed relative to comparable open-weights peers, the report treats Hermes 4 405B as a fully self-hostable option, useful where data residency rules out closed-weight APIs and where a 131K-token context supports long-document processing in a single pass.

Hermes 4 405B is a Nous Research open-weights fine-tune of Meta's Llama-3.1-405B, released in August 2025. According to the Unsloth re-hosted model card, training scaled the post-training corpus from about 1M samples to roughly 5M samples / 60B tokens blended across reasoning and non-reasoning data, and the model introduces a hybrid reasoning mode that can emit ... traces when the model decides to deliberate. The same page reports schema-adherent JSON output and a new RefusalBench where Hermes 4 claims SOTA among open and closed models for reduced refusal rates. The Unsloth model card also notes use of the Llama-3-Chat format with role headers and the ``thinking=True`` flag to activate reasoning, alongside an Unsloth Dynamic 2.0 GGUF quantization track optimized for llama.cpp via ``--jinja``. The mission framing emphasizes open, steerable behavior with broadly neutral alignment rather than heavy censorship. Independent confirmer sources place the model at about 406B parameters with a roughly 131K-token context window, consistent with this card.