A developer I know described opening the Hermes 4 technical report at 8 a.m. and spending the next hour wondering why their hosted deployment scored 20 points lower than the headline number. The answer was in the fine print: the 96.3% MATH-500 score requires reasoning mode on, and most API providers serve the model with it off by default.
That gap is the real story of the Hermes 4 release. Everything else-the open weights, the low price, the RefusalBench results-flows from understanding it.
What Hermes 4 actually released
Hermes 4 is a family of open-weight, hybrid-reasoning LLMs from Nous Research, released August 26, 2025, built on Meta's Llama 3.1 in 14B, 70B, and 405B sizes. The lab simultaneously pushed weights to Hugging Face and published a 94-page technical report.
The headline feature is what Nous Research calls "hybrid reasoning," which lets users 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.
The models are trained using a modified TorchTitan stack, starting from Llama 3.1 (405B, 70B) and Qwen3 14B checkpoints. The training regime incorporates loss-masking, length-control fine-tuning, and efficient sample packing, achieving >99.9% batch efficiency.
Training was conducted on 192 NVIDIA B200 GPUs, with a mixture of DDP, TP, and FSDP depending on model size.
The toggle is the product, not the benchmark
Hermes 4 405B is not "Llama 3.1 but better at math." It is a single checkpoint that switches between standard inference and chain-of-thought reasoning via a <think> toggle-and the 96.3% MATH-500 score is the reasoning-on number, which most hosted deployments do not enable by default.
Nous Research claims hybrid mode raises MATH-500 accuracy from 93.1% to 96.3% on the 405B tier. That three-point delta is real but the implication is easy to miss: every percentage point of that improvement costs you tokens. Reasoning mode expands outputs substantially-internal chains can run thousands of tokens before the model produces a final answer.
This creates a concrete production decision. The toggle allows users to request explicit chain-of-thought enclosed within tags. Developers can disable traces and restore faster, cheaper inference during production calls. The right pattern is to enable reasoning on hard math, coding, or logic tasks and leave it off for structured extraction, summarization, and retrieval-not to pick one mode globally.
The model offers major improvements in math, code, STEM, logic, and creative writing while supporting JSON schema adherence, function calling, and reduced refusal rates compared to other models.
Analysts caution that RefusalBench originates from Nous, not an independent lab.
The model scored 57.1% on RefusalBench, significantly outperforming GPT-4o (17.67%) and Claude Sonnet 4 (17%). That result is interesting, but the methodology is worth reading before building policy around it.
Hermes 4.3: the decentralized training experiment
The more technically interesting release came in December 2025. Hermes 4.3 36B is based on ByteDance's Seed 36B architecture and is notable for being the first Hermes model trained in a decentralized manner over the internet using the Psyche network.
Psyche is Nous Research's decentralized AI training orchestration platform, built on the Solana blockchain. The network applies DisTrO technology to enable collaborative model training across underutilized hardware distributed around the world.
Hermes 4.3 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 70B at half the parameter cost. Benchmark scores from the model card are concrete: MATH-500 at 93.8%, MMLU at 87.7%, BBH at 86.4%, AIME '24 at 71.9%, GPQA Diamond at 65.5%.
The Psyche-trained version even outperformed its traditionally centralized counterpart. That is the non-obvious result. The common assumption is that decentralized training introduces noise and inconsistency from heterogeneous hardware. Hermes 4.3 is early evidence that this assumption may not hold at mid-scale post-training runs, though a single data point is not a proof.
Available quantizations: GGUF variants at 4, 5, 6, and 8-bit are on HuggingFace under NousResearch/Hermes-4.3-36B-GGUF.
Based on Seed-OSS-36B-Base, Hermes 4.3 is an excellent shape for consumer local inference or enterprise self-deployment. The GGUFs comfortably sit in the VRAM of off-the-shelf GPUs.
Beagle in action#engineering, Tuesday 10:02amWhat the pricing arithmetic actually means for teams
Hermes 4 405B is priced at $1 per million input tokens and $3 per million output tokens, with a 131,072 token context window via OpenRouter. That is the list rate; actual clearing prices can run lower depending on provider competition.
To make this concrete: a moderately complex agent task consuming 8,000 input tokens and 1,500 output tokens costs roughly $0.012 per call at Hermes 4 405B rates. At GPT-4o pricing (~$2.50 per million input tokens, $10 per million output tokens as of 2025), a moderately complex agent task might cost $0.10-$0.50 per run. Run ten thousand agent tasks a month and that differential is real budget.
The catch is well-documented. Open-weight models dominate the low- and mid-cost portion of the performance frontier, but the lowest-error end is largely defined by closed-weight models, suggesting proprietary systems disproportionately push peak benchmark performance currently attainable.
| Dimension | Hermes 4 405B | Closed frontier (GPT-4o class) |
|---|---|---|
| Input price (per 1M tokens) | $1.00 | $2.50-$15.00 |
| Output price (per 1M tokens) | $3.00 | $10.00-$75.00 |
| MATH-500 (reasoning on) | 96.3% | Comparable or higher |
| GPQA Diamond | 70.5% | Often higher |
| Context window | 131K | 128K-200K |
| Can self-host | Yes | No |
| Version-pin stable | Yes | Depends on provider |
Model routing is the practical answer. Most production agent stacks benefit from using cheaper open-weight models for routine steps and routing complex reasoning to frontier models selectively.
A teammate like Beagle would fit naturally into that routing pattern-handling the lookup, extraction, and summarization steps at open-weight cost while escalating the genuinely hard reasoning calls.
Hermes 4 open-weight model: common questions
What is Hermes 4 and who made it?
Hermes 4 is an open-weight model family from Nous Research, released August 26, 2025. It runs in 14B, 70B, and 405B parameter sizes, built on Meta's Llama 3.1 and Qwen3 base models. The weights are public on Hugging Face. It is not a product from Meta or OpenAI.
What is the hybrid reasoning toggle in Hermes 4?
Hermes 4 centers on hybrid reasoning, allowing either fast direct responses or structured reasoning traces within <think> tags, giving users control over speed-accuracy tradeoffs in a way similar in spirit to reasoning-mode systems but with transparent configuration and behavior.
Enabling it improves hard math and coding accuracy; disabling it reduces token cost and latency.
How does Hermes 4 pricing compare to GPT-4o?
Hermes 4 405B is priced at $1 per million input tokens and $3 per million output tokens via OpenRouter. GPT-4o runs at $2.50 per million input and $10 per million output. For high-volume agent workloads, the difference compounds quickly-though closed frontier models still lead on the hardest reasoning tasks.
What is Hermes 4.3 and how is it different from Hermes 4?
Hermes 4.3 supports an extended context length of up to 512K tokens and delivers performance roughly equivalent to Hermes 4 70B at half the parameter count. It was also the first Hermes model trained entirely on Nous Research's Psyche decentralized network rather than a traditional GPU cluster-making it the more interesting research result, independent of the benchmark numbers.
Should a team use Hermes 4 or stick with a closed model?
Use open-weight models for control, version pinning, and self-hosted compliance. Prefer closed frontier models for the toughest reasoning and complex agentic workflows. Test models on representative workloads; benchmarks miss reliability and retry costs. The Hermes 4 pricing makes a hybrid routing strategy the most defensible default for most teams.