Frontier inference pricing has doubled since January 2026 alone. Meanwhile, a 36B open-weight model trained across commodity GPUs over the internet - no centralized cluster, no nine-figure compute bill - is hitting MATH-500 scores that beat its larger 70B sibling. That is the Hermes 4.3 situation in two sentences, and it is weirder and more consequential than the usual open-weight release.
What Hermes 4.3 actually is
Hermes 4.3 is the most technically interesting release in the 2025-2026 window, released August 25, 2025, built on ByteDance's Seed 36B base model. The naming is layered enough to cause confusion, so let's be precise: Hermes 4 is the language model. Hermes Agent is the self-improving agent framework that runs on top of a model. Hermes 4.3 is a new version of the former - not the agent, not the desktop GUI that shipped in June.
The model itself has three things worth paying attention to:
- Post-training corpus size. The dataset expanded from 1M samples and 1.2B tokens to roughly 5M samples and 60B tokens, blended across reasoning and non-reasoning data. That is a 50× increase in tokens. It shows in the benchmark results.
- Hybrid reasoning mode. The model can emit explicit thinking segments with special tags for chain-of-thought style internal reasoning while still delivering concise final answers. This is switchable - you do not pay the token cost when you do not need it.
- RefusalBench. Hermes 4.3 scored 74.6% on RefusalBench - meaning it answered 74.6% of questions that other aligned models refuse - compared to 59.5% for Hermes 4 70B. For teams where over-refusal on legitimate business tasks has caused workflow failures, this is a concrete and measurable improvement.
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%.
These are strong scores for a 36B model and outperform the larger Hermes 4 70B on several benchmarks according to the model card. The parameter-efficiency story is real, not marketing.
What Psyche and DisTrO actually did
This is the genuinely new part, and most coverage buries it.
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 out through data centers over the open internet, secured by the consensus of the Solana blockchain.
The engineering problem with training over the internet has always been bandwidth. Gradient updates between nodes in a traditional cluster travel over high-speed interconnects (InfiniBand, NVLink). Do it over commodity broadband and you either saturate the pipe or wait forever. DisTrO's answer:
- DisTrO reduces the amount of data exchanged between nodes during training by several orders of magnitude, enabling training over standard broadband connections. The concept is similar to image compression: much of the essential information in a model's gradient can be retained by transmitting only a few low-frequency components.
- DisTrO goes further by transmitting just the sign of each frequency amplitude, quantizing it down to one bit - resulting in roughly a 3× further reduction in data transmission.
- In large-scale tests, DisTrO achieved a bandwidth reduction factor of 1,000× to 10,000× compared to conventional solutions, eliminating costly hardware interconnections while maintaining high performance.
The result that nobody predicted: the Psyche-trained version of Hermes 4.3 outperformed the traditional centralized version on a suite of downstream tasks - a confirming signal that Psyche is up to the task of training production models.
That last sentence deserves a pause. The hypothesis going into the experiment was "decentralized training will probably match centralized." The actual outcome was that it beat it. Nous Research published both versions as a research artifact so the comparison is checkable.
The cost context that makes this matter right now
Decentralized training is interesting on its own. It gets more interesting when you look at what is happening to the alternative.
The AI inference market has split: the floor is collapsing while the ceiling is rising. Both things are true simultaneously, and confusing the two is what causes most enterprise AI budgets to blow up.
The cheapest API achieving GPT-4-class benchmark quality costs $0.14 per million input tokens as of late July 2026 (DeepSeek V4-Flash), versus $30.00 at GPT-4's March 2023 launch - a 214× reduction in 40 months. That is the collapsing floor. But the frontier ceiling is a different story: on April 23, GPT-5.5 launched at $5/$30 per million tokens, doubling GPT-5.4 on both sides.
A new Anthropic tokenizer produced up to 35% more tokens per prompt, so the effective rate climbed accordingly even when the sticker price stayed flat.
The number that puts all of this in concrete terms: at $0.13 input and $0.40 output per million tokens, Hermes 4 70B is inexpensive for a 70B open-weight model
- roughly 38× cheaper on input than GPT-5.5. Hermes 4.3's 36B size means it runs faster and on less hardware, which compounds that cost advantage further for self-hosted teams.
The honest implication: using frontier models for every task wastes money. Building classification logic that routes requests appropriately pays dividends. Hermes 4.3 is a strong candidate for the mid-tier route - the structured-output tasks, the reasoning workflows that do not need GPT-5.5-class capability, the self-hosted scenarios where data residency matters.
Beagle in action#ops-ai-review, weekly model cost auditWhat to watch and what to ignore
Watch: The Psyche decentralized training network is clearly an active investment. Hermes 4.3 being the first model trained this way suggests Nous Research is treating Psyche as a production training infrastructure path, not a one-off experiment. If decentralized training at production quality becomes repeatable, it removes the compute moat that concentrates model development among a handful of well-funded labs.
Watch: Nous Research is finalizing a new round of funding led by Robot Ventures, with participation from USV and other investors, at a $1.5 billion valuation - raising at least $75 million. That is a meaningful jump from the $1B valuation at which they raised from Paradigm. The money will likely go into Psyche infrastructure and the Hermes Agent ecosystem.
Ignore (for now): The blockchain layer. Solana handles coordination and fault-tolerance for Psyche nodes, but it is plumbing. Psyche reduces data transfer by several orders of magnitude, making distributed training practical; coordination happens on the Solana blockchain, ensuring a fault-tolerant and censorship-resistant network. Teams evaluating Hermes 4.3 as a model do not need to think about this at all. It is infrastructure concern, not a product one.
Be honest about: The benchmarks Nous Research uses to showcase RefusalBench are benchmarks Nous Research created. They acknowledge that standard math and coding benchmarks are easily gamed and are aiming for broader user agency
- which is fair - but RefusalBench is their own leaderboard. Run your own evals on your actual task distribution before shipping it on anything sensitive.
Hermes 4.3 open-weight model: common questions
What is Hermes 4.3 and how does it differ from Hermes 4?
Hermes 4.3 is a 36B open-weight model from Nous Research built on the ByteDance Seed base. It differs from Hermes 4 (70B) in two key ways: it was post-trained on a 50× larger dataset (60B tokens vs 1.2B), and it is the first Hermes model trained on Nous Research's decentralized Psyche network. Despite fewer parameters, it outperforms Hermes 4 70B on several standard benchmarks.
What is Nous Research's Psyche network?
Psyche is a decentralized AI training network that lets GPU nodes worldwide collaborate on a single training run over standard internet connections. It uses the DisTrO optimizer to reduce inter-node bandwidth by up to 10,000× compared to conventional gradient synchronization, making distributed training practical without high-speed cluster interconnects.
Is decentralized AI training production-ready?
Based on the Hermes 4.3 result, post-training on a decentralized network is production-ready at the 36B scale. Nous Research ran parallel centralized and decentralized training runs; the Psyche-trained version outperformed the centralized one on downstream tasks. Pre-training a full model from scratch at frontier scale over the internet remains an open research question.
When does it make sense to use Hermes 4.3 over a closed frontier model?
When your task is structured reasoning, JSON output, or instruction-following work that does not require the top-tier agentic coding or long-context capability of GPT-5.5 or Claude Fable 5 - and when data residency, cost predictability, or over-refusal from default-aligned models is a concern. At $0.13/$0.40 per million tokens via OpenRouter, the cost comparison is stark for high-volume workloads.
What does Hermes 4.3's RefusalBench score actually mean?
RefusalBench measures how often a model completes requests that standard aligned models refuse. Hermes 4.3 scores 74.6%, up from 59.5% for Hermes 4 70B. This matters for business tasks where conservative system-level refusals block legitimate workflows. Caveat: RefusalBench is Nous Research's own benchmark. Treat the number as directional, not definitive, and validate against your real task distribution.