A developer I know spent an afternoon last month auditing why her agentic pipeline kept stalling mid-task. The culprit was not a bug. The model was refusing a step it had silently decided was "sensitive" - a routine file-rename operation flagged because the path contained the word "delete." She swapped to a Hermes model that afternoon and hasn't looked back.
That small scene captures what makes the Hermes open-weight model line worth understanding. It is not trying to be the biggest or the fastest. It is trying to be the one that actually does what you told it to do.
What changed between Hermes 3 and Hermes 4
Nous Research has released three generations since August 2024, each fine-tuned on an open-weight base and built to fix a specific failure mode in the prior version. The short version: Hermes 3 added structured function calling, Hermes 4 added chain-of-thought reasoning, and Hermes 4.3 solved the over-refusal problem while extending context to 512K tokens.
Hermes 4, released August 2025, is a family of open-weight models - 14B, 70B, and 405B parameter sizes, based on Llama 3.1 checkpoints - that achieves its performance entirely through post-training techniques. No novel architecture. No proprietary data moat. Just aggressive post-training, done in public, with weights you can download.
Hermes 4 introduces hybrid reasoning: models can toggle between standard responses and explicit reasoning using <think>...</think> tags when complex problems require deeper deliberation. When activated, the models generate their internal reasoning within those tags before providing a final answer - similar to OpenAI's o1 models but with full transparency into the thinking process.
The benchmarks on the 405B variant are strong. It 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 numbers are competitive with proprietary frontier models. The honest caveat: the 405B variant is a multi-GPU proposition. Running a 405B model requires either a multi-GPU cluster or cloud inference. Most production deployments use the 70B or smaller variants for cost and latency reasons.
The training infrastructure is worth noting too. Hermes 4 uses Atropos, Nous Research's open-source reinforcement learning environment, to implement rejection sampling across approximately 1,000 different task-specific verifiers - filtering for high-quality reasoning trajectories across diverse domains.
Key verification environments include Answer Format Training covering 150+ output formats, Instruction Following, Schema Adherence for JSON generation using Pydantic models, and Tool Use training for agentic behavior. That last item is the one teams building pipelines should pay attention to.
There was one real problem Nous found during development. One of Hermes 4's most significant challenges was the overlong reasoning problem - where reasoning models generate excessively long chains of thought without termination. The research team discovered their 14B model reached maximum context length 60% of the time on LiveCodeBench when in reasoning mode. This is not a Hermes-only problem - it affects most reasoning models - but the fact that Nous published it openly in their technical report is the kind of transparency that makes the Hermes line easier to evaluate honestly.
Why Hermes 4.3 is the release that matters for agents
Hermes 4.3 is a frontier, hybrid-mode reasoning model based on the ByteDance Seed 36B base. The move away from Llama is notable - it signals that Nous is no longer constrained to Meta's release cadence, and the Seed 36B base has strong coding and reasoning priors that complement the Hermes post-training approach.
The number that stands out for anyone running automated workflows is RefusalBench. The defining result was RefusalBench: 74.6% versus GPT-4o's 17.67%. RefusalBench measures whether a model follows user constraints precisely - whether it does what you said, in the way you specified, without adding unsolicited caveats or refusing legitimate requests. For automated agent workflows, this is not a secondary metric. When a model refuses or softens an instruction mid-pipeline, the automation logic breaks - sometimes silently. A 74.6% versus 17.67% gap is not a small margin.
The post-training corpus for Hermes 4.3 grew from 1M samples and 1.2B tokens to approximately 5M samples and 60B tokens, blended across reasoning and non-reasoning data. That 50x expansion in token count, applied to a well-chosen base model, is what produces the step-change on instruction-following.
The Psyche training story - genuinely new, still worth watching
This is the part of the Hermes 4.3 release that is genuinely unprecedented, and also the part where some hype-checking is warranted.
Hermes 4.3 is Nous Research's first production model post-trained entirely on the Psyche network - a distributed training system that uses the DisTrO optimizer to efficiently communicate between training nodes spread across data centers over the open internet, secured by the consensus of the Solana blockchain.
Nous verified its effectiveness for production workloads by training Hermes 4.3 both on Psyche and via the traditional centralized approach using a custom version of Torchitan, then trained the model a second time on Psyche using tensor parallelism and DisTrO.
The training run proved stable, averaging 144,000 tokens per second across 24 Psyche nodes. Using DisTrO's overlapped collective strategy, the P2P communications were hidden by the training time, effectively achieving equivalent throughput to traditional centralized training.
The Psyche-trained version of Hermes 4.3 outperformed the traditional centralized version. That is the detail worth sitting with. Decentralized training is not new as a research idea, but producing a model that is actually competitive - and doing it openly, with verifiable on-chain records - is new as a production outcome.
The honest caveat: the Psyche decentralized training network is clearly an active investment, and Hermes 4.3 being explicitly the first model trained this way suggests Nous Research is treating Psyche as a production training infrastructure path, not a one-off experiment. Whether the approach scales to pre-training at the 100B+ parameter range remains unproven. Consilience 40B has reached 20 trillion training tokens via Psyche , which suggests the pre-training path is being actively pursued, but that model is not yet publicly benchmarked against the current field.
What teams should actually take from this
At $0.13 input and $0.40 output per million tokens, Hermes 4 70B pricing is competitive for a 70B-class model, and the 131K context window is standard for this class. For teams interested in open-weight fine-tuning or self-hosted deployment, it is worth evaluating.
For teams self-hosting: you need a local inference backend (Ollama is the simplest path), a local model (Hermes 4.3 GGUF variants work well), and you need to disable or not configure any cloud provider endpoints if you are operating under compliance requirements like HIPAA or CMMC.
The Hermes line is not a replacement for every use case. Math-heavy reasoning at scale still favors the 405B variant or closed frontier models with more training compute behind them. But for teams running structured agent workflows - tool calls, JSON schema adherence, multi-step instruction following - Hermes 4.3 is the most honest open-weight option currently available at the 36B class. It does what you ask. That turns out to be harder than it sounds.
Beagle connects to Slack and Microsoft Teams; for the teams we work with, the model they route agent tasks through matters more than the wrapper. If you are evaluating Hermes for an agentic workflow, the use cases section has context on where structured tool-calling models tend to outperform general assistants in practice.