# Hermes 3 70B ![Organization](https://img.shields.io/badge/Org-NousResearch-purple) ![Parameters](https://img.shields.io/badge/Params-70B-1f6feb) ![License](https://img.shields.io/badge/License-Llama%203.1-d29922) ![Weights](https://img.shields.io/badge/Weights-Open-2da44e) ![Released](https://img.shields.io/badge/Released-Aug%202024-555) > **Quick answer:** Hermes 3 70B is NousResearch's August 2024 instruction fine-tune of Meta's Llama 3.1 70B. It scores 47.2% on MMLU-Pro. Under the Llama 3.1 License with open weights on Hugging Face, it is designed for strong general reasoning, long-context tasks, and agentic deployment patterns. ## At a Glance **Where Hermes 3 70B leads** - NousResearch instruction-tuning — improved function calling and tool use over base Llama 3.1 70B - 128K context window — full Llama 3.1 70B context - Open weights — self-hostable for agentic pipelines - Strong at complex system prompts and multi-turn conversation - Agent-optimised: designed explicitly for agentic and tool-using deployment **Where it lags** - 47.2% MMLU-Pro — below leading 2025 open-weight models (Qwen3 32B, Llama 4 Maverick) - Base Llama 3.1 70B from August 2024 — older than Llama 4, Qwen3, Phi-4 - Knowledge cutoff March 2024 — outdated - No reasoning/thinking mode **Best for:** Agentic and tool-use pipelines requiring open-weight models; teams specifically using NousResearch's Hermes instruction style; legacy Hermes-based deployments. ## What Hermes 3 70B Is Hermes 3 70B is NousResearch's third generation Hermes fine-tune, built on Meta's Llama 3.1 70B base. Released August 12, 2024, it specialises in advanced instruction following, function calling, and agentic capability — areas where the base Llama 3.1 70B was strong but could be improved through targeted fine-tuning. The Hermes series is known in the open-source community for producing models with reliable tool use, structured output generation, and consistent behaviour with complex system prompts. Hermes 3 70B brings these properties to the Llama 3.1 70B scale. Its 47.2% MMLU-Pro score reflects its 2024 vintage and 70B base — newer open-weight models like Llama 4 Maverick (80.5%), Qwen3 32B (no MMLU-Pro score but strong AetherCode), and Phi-4 (70.4% at only 14B) have since surpassed it on academic knowledge benchmarks. Hermes 3 70B remains relevant for use cases that specifically benefit from NousResearch's instruction-tuning approach and are already integrated with Hermes-style prompting. ## Specifications | Field | Value | |---|---| | Organization | NousResearch | | Base model | Meta Llama 3.1 70B | | Parameters | 70B (dense) | | Context window | 128,000 tokens | | License | Llama 3.1 License | | HuggingFace | [NousResearch/Hermes-3-Llama-3.1-70B](https://huggingface.co/NousResearch/Hermes-3-Llama-3.1-70B) | | Release date | August 12, 2024 | | Knowledge cutoff | March 2024 | | Modality | Text only | | Fine-tune type | Instruction + agentic | ## Pricing Hermes 3 70B is available as open weights under the Llama 3.1 License — free to self-host. Available via third-party providers (Together AI, Replicate, etc.) at market rates. ## Context Window Hermes 3 70B has a **128,000-token context window** — the full Llama 3.1 70B context, suitable for long document analysis and multi-turn agentic tasks. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [MMLU-Pro](/benchmarks/mmlu-pro) | 47.2% | Benchgen evaluation | 2025-07 | ## Hermes 3 70B vs Alternatives | Model | MMLU-Pro | Context | License | Type | |---|---|---|---|---| | **Hermes 3 70B** | 47.2% | 128K | Llama 3.1 | Instruction FT | | [Llama 4 Maverick](/models/meta/llama-4-maverick) | 80.5% | 1M | Llama 4 | Native | | [Phi-4](/models/microsoft/phi-4) | 70.4% | 16K | Apache 2.0 | Native | | [Qwen3 32B](/models/qwen/qwen3-32b) | — | 131K | Apache 2.0 | Native | For new deployments requiring open-weight general capability, Llama 4 Maverick, Qwen3 32B, or Phi-4 provide significantly better benchmark performance. Hermes 3 70B is most relevant when the specific Hermes instruction-tuning style or existing Hermes-based production integrations are the primary consideration. ## Related Models & Benchmarks - Compare: [Llama 4 Maverick](/models/meta/llama-4-maverick) · [Phi-4](/models/microsoft/phi-4) · [Qwen3 32B](/models/qwen/qwen3-32b) - Benchmarks: [MMLU-Pro](/benchmarks/mmlu-pro) ## Frequently Asked Questions

What is Hermes 3 70B?

Hermes 3 70B is NousResearch's August 2024 instruction fine-tune of Llama 3.1 70B, scoring 47.2% MMLU-Pro. It is optimised for agentic tasks, function calling, and complex instruction following.

Is Hermes 3 70B open source?

Yes. Hermes 3 70B is available on Hugging Face under the Llama 3.1 License — permitting commercial use within the Llama 3.1 license terms.

What is the difference between Hermes 3 and the base Llama 3.1?

Hermes 3 adds NousResearch's instruction fine-tuning on top of Llama 3.1 70B, improving function calling, tool use, structured output generation, and complex system prompt handling beyond the base model.

--- *Specs from NousResearch's official Hermes 3 release (August 2024) and Benchgen evaluations. Last updated 2026-07-24.*

Hermes 3 70B — Evaluations

Evaluations