Hermes 4's largest 405B model scored 96.3% on MATH-500 and 81.9% on AIME '24 in reasoning mode. That number landed in every launch headline on August 26, 2025. Most of those headlines missed the point.
The interesting thing is not the score. It is how the score was achieved, and what it means for teams who actually want to run an open-weight reasoning model without managing two separate checkpoints.
What Hermes 4 actually ships
Hermes 4 is a family of hybrid reasoning models that combine structured, multi-turn reasoning with broad instruction-following ability.
The family includes a flagship 405B parameter variant based on Llama-3.1-405B, with smaller 14B and 70B versions for lighter deployments, all released under an open philosophy with downloadable weights.
The 14B sits on Qwen 3 14B base and carries an Apache 2.0 license. That matters for teams with commercial use requirements - the 405B is under the more restrictive Llama 3 license.
Hermes 4 introduces what Nous Research calls "hybrid reasoning," allowing users to toggle between fast responses and deeper, step-by-step thinking. 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.
That toggle is the genuinely new part. Most reasoning models ship as separate checkpoints from their base: you pick the reasoning variant or the chat variant and live with the tradeoff. Hermes 4 puts both on the same weights, controlled at inference time with a boolean flag. For a team building an agent that sometimes needs fast responses and sometimes needs to grind through a multi-step problem, that is a meaningful operational simplification.
The post-training data story is also more substantial than it looks. The Hermes 4 dataset consists primarily of newly synthesized reasoning and non-reasoning data, totaling approximately 5 million samples and 19 billion tokens.
The post-training corpus for Hermes 4 was approximately 1M samples covering around 1.2 billion tokens, according to the Hermes 4.3 model card which compares its own expanded corpus against the Hermes 4 baseline
- so the 405B's corpus roughly 15x the sample count from the prior generation.
Where the benchmark framing misleads
The 96.3% figure is real. It comes directly from the technical report published alongside the weights (arXiv 2508.18255). On AIME 2024, Hermes 4 405B scores 81.9%. Those are strong numbers. The problem is the comparison.
Hermes 4 405B (non-reasoning mode) scores 9 on the Artificial Analysis Intelligence Index, placing it below average among other open-weight non-reasoning models of similar size, where the median is 18. It generates output at 36.6 tokens per second, which is at the lower end compared to models of similar size (median: 60.2 t/s).
Pricing via hosted providers sits at $1.00 per million input tokens and $3.00 per million output tokens - well above the median for open-weight models of similar size.
So: strong task-specific math scores, below-median composite intelligence ranking, below-median inference speed, above-median price. Nous positioned the 405B as "competitive with DeepSeek R1 and Qwen3 235B." The community consensus was that DeepSeek R1 at 671B outperforms Hermes 4 on raw benchmarks, and Qwen3 235B is in the same conversation. That is a fair pushback. MATH-500 is a clean, verifiable benchmark - but composite rankings that include instruction following, coding, and real-world tasks tell a more complete story.
The honest read: Hermes 4 is a strong fine-tune of capable base models, competitive in math and reasoning, average-to-below on broader capability measures, and interesting specifically for the alignment stance and the toggle mechanic. It is not a generational leap past DeepSeek or Qwen at the raw capability level.
Hermes 4.3 and decentralized training via Psyche
This is where the release gets genuinely structurally interesting, and it received almost no coverage relative to the MATH score.
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, with consensus secured by the Solana blockchain.
The claim sounds experimental. The results suggest it is not. The training run averaged 144,000 tokens per second spread across 24 Psyche nodes. Using DisTrO's overlapped collective strategy, the entirety of the P2P communications were hidden by the training time, effectively achieving equivalent throughput to traditional centralized training.
Nous verified the approach by training Hermes 4.3 both on Psyche and via the traditional centralized approach , then compared results. The Psyche-trained version outperformed the traditional centralized version.
Hermes 4.3 36B is a hybrid-mode reasoning model based on ByteDance's Seed 36B base.
It was trained with an extended context length of up to 512K tokens and nearly matches - and in some cases exceeds - the performance of Hermes 4.
The post-training corpus for Hermes 4.3 expanded to approximately 5 million samples covering around 60 billion tokens, compared to roughly 1.2 billion tokens for Hermes 4 - a roughly 50x expansion in training token count for the fine-tuning stage.
The decentralized training angle matters beyond Nous Research. If you can run competitive post-training across geographically distributed nodes at parity with centralized compute, the hardware moat that currently concentrates AI development at a handful of labs starts to look more porous. That is a second-order consequence worth watching.
Beagle in action#ml-infra, 11:42amHow Hermes 4 fits into the open-weight landscape
The broader context makes Hermes 4 easier to place. Frontier open-weight models lag behind the most capable closed models by an average of three months in the Epoch Capabilities Index.
The gap varies considerably over time, sometimes even closing completely.
On many common enterprise tasks - coding, text classification, summarization, structured data extraction, instruction following - the best open-weight models now perform comparably to GPT-4o and Claude Sonnet. On the most complex reasoning tasks and in long agentic workflows, closed frontier models still hold an edge.
Hermes 4 sits in a specific part of that landscape:
| Dimension | Hermes 4 14B | Hermes 4 405B | Hermes 4.3 36B |
|---|---|---|---|
| Base model | Qwen 3 14B | Llama 3.1 405B | ByteDance Seed 36B |
| License | Apache 2.0 | Llama 3 | Seed 36B license |
| Context window | 128K | 128K | 512K |
| Reasoning toggle | Yes | Yes | Yes |
| Training method | Centralized | Centralized | Decentralized (Psyche) |
| Best fit | Local / commercial use | High-math workloads, hosted | Long-context, experimental infra |
The alignment stance - minimal refusals, user-steerable behavior - is a deliberate product decision, not an oversight. Hermes 4 is designed with minimal built-in content filters or refusals. This neutrally aligned approach aims to give users greater control, enabling the model to engage with a wide range of queries without excessive censorship. Teams in regulated industries should treat that as a signal to add their own guardrails at the application layer rather than relying on the model.
A teammate like Beagle, running Hermes 4 locally via a tool-call interface, would be operating against a model with no platform-level content policy - which means your review step before any send matters more, not less.
Hermes 4 open weight model: common questions
What is Hermes 4 and who made it?
Hermes 4 is a family of open-weight hybrid reasoning models released by Nous Research on August 26, 2025. It comes in 14B, 70B, and 405B parameter sizes, built on Llama 3.1 and Qwen 3 base models. The defining feature is a per-request reasoning toggle on a single checkpoint, with no separate model required.
How does Hermes 4's hybrid reasoning mode work?
When reasoning is enabled, the model emits a <think>...</think> block containing its chain-of-thought before the final answer. When disabled, it responds directly. Both behaviors live on the same weights, controlled at inference time with a boolean parameter. This avoids the need to manage two separate model deployments.
Is Hermes 4 actually competitive with DeepSeek R1?
On MATH-500 and AIME benchmarks, the 405B variant is in the same range. On broader composite evaluations - Artificial Analysis Intelligence Index, instruction following, multi-step agentic tasks - Hermes 4 405B scores below the median for its size class. The comparison holds on specific math tasks; it does not hold generally.
What is Hermes 4.3 and how is it different?
Hermes 4.3 is a 36B model based on ByteDance's Seed 36B, released in December 2025. Its defining difference is that it was post-trained entirely on Nous Research's Psyche decentralized network - the first Hermes model trained this way. It also extends the context window to 512K tokens and expanded the post-training corpus to roughly 60 billion tokens.
Which Hermes 4 variant should a team actually use?
The 14B is the practical starting point: Apache 2.0 licensed, fits on a single modern GPU, and the reasoning toggle works. The 4.3 36B is worth testing if you need longer context or want to track the decentralized training experiment. The 405B makes sense only if hosted inference cost is not a concern and the workload is genuinely math-heavy.