Hermes 4 is a hybrid-reasoning model family trained on 5 million post-training samples across 19 billion tokens - 3.5 million of them reasoning-heavy examples with sequences up to 16K tokens - using DataForge for synthetic data generation and Atropos for rejection sampling across task-specific RL environments. On AIME'24 it scores 81.9% and 61.3% on LiveCodeBench, outperforming DeepSeek-R1, while cutting overlong outputs by 78%. The headline claim is not about a new base model. It is about what post-training alone can do to a Llama 3.1 checkpoint.
That distinction matters more than most coverage acknowledges.
What Hermes 4 actually is (and what it is not)
Hermes 4 is a family of hybrid reasoning models that combine structured, multi-turn reasoning with broad instruction-following ability.
The key word is hybrid.
Models can toggle between standard responses and explicit reasoning using <think>...</think> tags when complex problems require deeper deliberation.
You control this per-request - fast mode for simple lookups, deliberation mode for code review or multi-step logic.
The family comes in 14B, 70B, and 405B parameter sizes based on Llama 3.1 checkpoints and achieves frontier-level performance through pure post-training techniques. Nothing about the base architecture changed. What changed is the training recipe. DataForge processes pre-training seed data through a graph-based synthetic data generator inspired by AgentInstruct, producing conversational data across a wide variety of tasks.
Atropos is an open-source reinforcement learning framework that provides hundreds of specialized training environments, allowing the model to practice specific skills.
The practical implication: if you wanted to replicate this approach for a domain-specific fine-tune, both tools are public. The recipe is open.
One thing the coverage mostly glosses over: the 14B variant is based on Qwen3-14B, not Llama 3.1. The 70B and 405B are Llama 3.1. That matters for licensing. The 14B ships under Apache 2.0. The 70B and 405B carry the Llama 3.1 community license, which restricts use at scale. If you are building a product serving more than 700 million monthly active users, you need a separate Meta license for the Llama variants. The 14B does not have that restriction.
The 405B versus the 70B: a cost question, not a quality question
Hermes 4 70B is priced at $0.13 per million input tokens and $0.40 per million output tokens, with a 131,072-token context window on OpenRouter.
The 405B costs $1.00 per million input tokens and $3.00 per million output tokens, also with a 131,072-token context window. That is roughly an 8x output cost difference.
For most agent workloads in Slack or a CI pipeline - where the task is structured (write a test, summarize a thread, classify a ticket) - the 70B is the correct starting point. The 405B earns its cost on tasks where deliberation depth pays off: complex multi-step code generation, hard math, or long-document synthesis where errors compound.
| Model | Input $/M | Output $/M | Base | License |
|---|---|---|---|---|
| Hermes 4 14B | (API varies) | (API varies) | Qwen3-14B | Apache 2.0 |
| Hermes 4 70B | $0.13 | $0.40 | Llama 3.1 70B | Llama 3.1 community |
| Hermes 4 405B | $1.00 | $3.00 | Llama 3.1 405B | Llama 3.1 community |
| Hermes 4.3 36B | (self-host/Nous Portal) | - | Seed-OSS 36B | Apache 2.0 |
Hermes 4.3 is the more interesting release
Hermes 4.3 was trained with a 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 designed as an excellent shape for consumer local inference or enterprise self-deployment.
What makes it worth understanding in depth is the training infrastructure. Hermes 4.3 is Nous Research's first production model post-trained entirely on the Psyche network, a distributed training network 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.
To verify the approach, Nous trained Hermes 4.3 both on Psyche and via the traditional centralized approach
- and compared the results directly.
Benchmark scores from the model card: MATH-500 at 93.8%, MMLU at 87.7%, BBH at 86.4%, AIME'24 at 71.9%, GPQA Diamond at 65.5%.
Despite having roughly half the parameters of Hermes 4 70B, it nearly matches or surpasses it on several benchmarks.
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. That claim is worth holding lightly for now. The training results are competitive, which proves viability. Whether decentralized training at this scale becomes a repeatable production method - or whether the coordination overhead limits how far it scales - is still an open question.
A teammate running inside Slack can call Hermes 4.3 at self-hosted rates for structured tasks (ticket classification, PR summaries, draft replies) without the per-seat cost of a hosted frontier service.
Beagle in action#engineering, mid-sprintWhere the hype ends
The benchmarks are real. The "outperforms ChatGPT" framing most coverage ran with is not wrong, but it is comparing against versions of GPT-4 that are no longer the current standard. Nous Research itself notes that standard math and coding benchmarks are easily gamed
- an honest caveat buried in the 4.3 launch post.
The alignment story also needs context. Unlike models from OpenAI, Google, or Anthropic, Hermes 4 is designed to respond to almost any request without the safety protections commonly found in commercial AI systems. For some teams - security research, red-teaming, creative applications - that is a feature. For enterprise deployments where you need documented refusal behavior and auditability, it is a gap your own guardrail layer has to fill.
While per-token inference prices generally decline, the cost of running frontier-level models on difficult benchmarks has increased at an approximately exponential rate. These trends are not contradictory: per-token inference can become cheaper while frontier performance rises in cost if achieving marginal gains requires substantially more inference. Hermes 4 sits in a good spot on that curve - close to frontier performance, priced well below it - but that gap will not stay fixed. As the closed labs continue pushing reasoning depth, the post-training efficiency advantage Nous is demonstrating here will face more pressure.
The genuinely non-obvious thing Hermes 4 demonstrates: post-training is now a first-class research discipline, not a wrapper around a base model. Nous took a Llama 3.1 checkpoint and, with a 5-million-sample training corpus and two open-source tools, produced benchmark results that overlap with closed models costing an order of magnitude more to run. That is the finding worth sitting with.
Hermes 4 open-weight model: common questions
What is Hermes 4 and who makes it?
Hermes 4 is an open-weight, hybrid-reasoning model family from Nous Research, released in August 2025. It comes in 14B, 70B, and 405B sizes. The 14B is built on Qwen3-14B; the larger variants use Llama 3.1 as the base. Weights are downloadable from Hugging Face and accessible via API on OpenRouter.
How does Hermes 4 hybrid reasoning work?
The model can toggle between direct responses and explicit chain-of-thought using <think>...</think> tags. You control this per-request via a boolean or system prompt. Direct mode is faster and cheaper; reasoning mode costs more output tokens but improves accuracy on math, code, and multi-step logic tasks.
What is the difference between Hermes 4 and Hermes 4.3?
Hermes 4 (14B/70B/405B) was released in August 2025 using centralized GPU training. Hermes 4.3 (36B, Seed base) was released in December 2025 and is the first Hermes model trained on Nous Research's Psyche decentralized network. It supports up to 512K context and nearly matches Hermes 4 70B performance at half the parameter count, under Apache 2.0 license.
Can I run Hermes 4 locally?
The 14B fits comfortably on a single high-end consumer GPU. The 70B requires a multi-GPU setup or cloud inference. Hermes 4.3 36B is explicitly designed for local deployment - GGUF quantized versions in 4, 5, 6, and 8-bit are available on Hugging Face and fit in the VRAM of off-the-shelf GPUs.
What does Hermes 4's "neutral alignment" mean for enterprise use?
Hermes 4 is trained for low refusal rates and user-directed behavior, which means it will comply with requests that commercial models typically decline. For enterprise deployments, this requires your own guardrail or content-filtering layer - Hermes 4 does not provide one by default. Teams that need documented refusal behavior for compliance purposes should account for that engineering cost.