A researcher at a 12-person biotech opens a Slack thread at 8am asking for a JSON schema validated against their internal Pydantic models. The answer comes back in under a minute, well-formed, sourced, no hedging paragraph about what the model cannot help with. That is the experience Nous Research has been optimizing toward for three years. Hermes 4, released August 26 2025, is the most concrete version of that bet yet - and one number in the launch coverage tells you almost everything about why it's worth reading past the headline scores.
What Hermes 4 actually is, technically
Hermes 4 is a family of open-weight models - 14B, 70B, and 405B parameter sizes - based on Llama 3.1 checkpoints, that achieves frontier-level performance through pure post-training techniques. No new base model, no architectural invention: Hermes 4 builds on Llama 3.1 with extensive post-training that blends verified reasoning traces and non-reasoning data to improve math, coding, schema adherence, and general assistant quality while maintaining neutral alignment and steerability.
The training stack is the interesting part. Hermes 4 uses Atropos, Nous Research's open-source reinforcement learning environment, to implement rejection sampling across approximately 1,000 different task-specific verifiers - a verification infrastructure that filters for high-quality reasoning trajectories across diverse domains.
Key verification environments include Answer Format Training (rewarding correct formatting across 150+ output formats), Instruction Following using RLVR-IFEval tasks with complex constraints, Schema Adherence for JSON generation using Pydantic models, and Tool Use training for agentic behavior.
Atropos is Nous Research's reinforcement-learning framework for training and aligning models. It is a model-training pipeline, not the Hermes agent runtime - it sits upstream, shaping the kinds of agentic models Hermes and others then run. That distinction matters for teams evaluating whether Hermes 4 is genuinely better at tool-calling tasks, or just louder about being so. The 1,000-verifier rejection sampling approach is the mechanism that produces structured output quality. It's not standard RLHF - it's a rollout handler that manages asynchronous coordination across potentially thousands of distributed workers, addressing the challenge of highly variable LLM generation times.
The 405B model was trained using a modified TorchTitan across 192 NVIDIA B200 GPUs, handling highly heterogeneous sample length distribution through efficient packing that achieved greater than 99.9% batch efficiency, with flex attention and loss masking where only assistant-role tokens contribute to cross-entropy loss.
The benchmark scores that need context
The numbers Nous led with are real.
The 405B model achieves 96.3% on MATH-500 in reasoning mode, 81.9% on AIME'24, 78.1% on AIME'25, 70.5% on GPQA Diamond, and 61.3% on LiveCodeBench.
Those are competitive with models trained from scratch on dedicated reasoning infrastructure. The reasoning toggle is the mechanism:
Hermes 4 introduces "hybrid reasoning," allowing users to toggle between fast responses and deeper, step-by-step thinking processes. When activated, the models generate their internal reasoning within special <think> tags before providing a final answer - similar to OpenAI's o1 reasoning models but with full transparency into the AI's thought process.
Artificial Analysis ranked the 405B number 22 of 37 models specifically because they benchmarked it in non-reasoning mode. That detail alone changes who should care about this model and why. The hybrid toggle is central to understanding where Hermes 4 sits - it's not uniformly strong, it's strong when reasoning is active.
Now the RefusalBench number. Hermes 4 achieved the highest score among all tested models on "RefusalBench," a benchmark created by Nous Research to assess how often AI systems decline to answer questions. In reasoning mode the model scored 57.1%, significantly outperforming GPT-4o at 17.67% and Claude Sonnet 4 at 17%. That spread generated most of the press coverage. Two things to hold alongside it:
First, RefusalBench comprises 166 handcrafted prompts across 32 categories tested with an LLM-as-judge procedure to measure when models decline to answer, with three categories inverted to reward appropriate refusals related to minor harm, exploitation/trafficking, and self-harm. Nous built the benchmark. Citing it as independent evidence of alignment quality requires some skepticism.
Second, a detail almost no coverage mentioned: the Hermes 70B actually edged out the 405B on RefusalBench at 59.5%. If minimal refusals is the goal, the smaller model scores higher - which makes the flagship number slightly less clean than it appears.
Hermes 4.3: the more interesting release
Hermes 4 launched with the bigger benchmark numbers. Hermes 4.3, released December 3 2025, is the more technically interesting event. Hermes 4.3 36B is a frontier, hybrid-mode reasoning model based on ByteDance's Seed 36B base. It is the first Hermes model trained in a decentralized manner over the internet using Psyche.
Hermes 4.3 is the first production model post-trained entirely on the Psyche network - Nous Research's distributed training network that uses the DisTrO optimizer to efficiently communicate between training nodes spread out through data centers over the open internet and secured by the consensus of the Solana blockchain. By enabling nodes throughout the world to collaborate on a single training run, Psyche can dramatically reduce the cost of training frontier-level models, leveling the playing field for open-source AI model developers.
Whether decentralized training at this scale becomes a repeatable production method or remains an experiment is worth watching. The training results are competitive, which at minimum demonstrates that the approach is viable.
The Psyche-trained version even outperformed its traditionally centralized counterpart
- a result Nous reported without much fanfare that deserves more attention than it received.
Hermes 4.3 was trained with an extended context length up to 512K tokens and nearly matches - and in some cases exceeds - the performance of Hermes 4 70B at half the parameter cost. Based on Seed-OSS-36B-Base, it is an excellent shape for consumer local inference or enterprise self-deployment. The GGUFs comfortably sit in the VRAM of off-the-shelf GPUs.
The post-training corpus expanded massively: from 1M samples and 1.2B tokens up to approximately 5M samples and 60B tokens, blended across reasoning and non-reasoning data. That 50x corpus expansion, paired with a 36B model that fits in consumer GPU VRAM, makes Hermes 4.3 the version most teams doing local or self-hosted inference should actually evaluate - not the flagship 405B.
Beagle in action#research-ops, 2:17pmWhere Hermes 4 actually fits for teams
The practical decision for a team evaluating open-weight models is which size to run and at what cost. Via OpenRouter, the 405B is priced at $1 per million input tokens and $3 per million output tokens with a 131K context window. For comparison, a team running 200 structured-output requests per day at ~2,000 tokens average output would spend roughly $1.20/day on the 405B via API - manageable, but the 70B or 4.3 variants running self-hosted will cost a fraction of that at scale.
Hermes 4 releases include BF16/FP8 and GGUF formats via Hugging Face, easing local and server deployments across model sizes, with the 14B and 70B variants positioned for edge or resource-constrained environments and the 405B aimed at high-end setups or hosted APIs.
The ecosystem includes structured output modes like JSON adherence, function calling, and tool use, complementing the hybrid reasoning capabilities for agents and production workflows.
The honest summary: Hermes 4's reasoning scores are real and its tool-calling training is genuinely sophisticated - 1,000 verifiers across 150+ output formats is not a marketing claim, it's described in detail in the technical report on arXiv (2508.18255). The "uncensored AI" framing in most coverage is both accurate and a little reductive. What Nous has actually built is a model that treats structured outputs and tool calling as first-class training objectives, with an alignment stance that shifts responsibility for safety boundaries to the deploying team. That is a coherent design philosophy. It just requires you to own the safety layer, not outsource it to the model.
Nous Research Hermes 4: common questions
What is Hermes 4 and how does it differ from Hermes 3?
Hermes 4 is an open-weight model family from Nous Research (14B, 70B, 405B) built on Llama 3.1 and trained with hybrid reasoning via the Atropos RL framework. The key additions over Hermes 3 are the toggleable <think> reasoning mode, 1,000 task-specific verifiers for rejection sampling, and dedicated tool-use training environments. The post-training corpus is also substantially larger.
Does Hermes 4 run locally?
Yes. All three sizes ship as GGUF quantizations on Hugging Face, compatible with Ollama, llama.cpp, and LM Studio. Hermes 4.3 36B is the strongest local option - it fits in consumer GPU VRAM, runs a 512K context window, and nearly matches Hermes 4 70B performance at half the parameter count.
What is the RefusalBench score, and should I trust it?
RefusalBench measures how often a model refuses user requests across 166 prompts in 32 categories. Hermes 4 scored 57.1%, versus 17.67% for GPT-4o. The caveat: Nous Research created the benchmark themselves. It is a useful signal about the model's alignment stance, but treat it as directional rather than independent evidence.
How does Hermes 4.3 differ from Hermes 4?
Hermes 4.3 is a 36B model based on ByteDance's Seed base rather than Llama 3.1, and it's the first Hermes model trained on Nous Research's Psyche decentralized network. Its post-training corpus expanded from 1.2B to 60B tokens. It supports a 512K context window and outperforms Hermes 4 70B on several benchmarks despite being a smaller model.
Is Hermes 4 good for structured output and tool calling?
Yes - structured outputs are a first-class training target. Atropos verification environments explicitly train JSON schema adherence with Pydantic models, function calling, and 150+ output formats. Teams building agents that need reliable tool use and structured responses have the most to gain from evaluating Hermes 4 against closed-weights alternatives.