A 36-billion-parameter model trained across 24 nodes spread over the open internet, coordinated by a blockchain, hit 93.8% on MATH-500. That's what Nous Research shipped in August 2025 with Hermes 4.3. The training story is the genuinely new thing here - and it's worth separating from the marketing.
What Hermes 4.3 actually is
Hermes 4.3 36B is a frontier hybrid-mode reasoning model based on ByteDance's Seed 36B base. That's the first break from the series' history: Hermes 4 represents the point where Nous Research moved from fine-tuning Meta's Llama models to working with a broader set of base models. Prior to this, every Hermes release was a Llama fine-tune.
The model runs in two modes.
It emits explicit thinking segments with special tags for deeper, chain-of-thought style internal reasoning while still delivering concise final answers.
You toggle this at inference time: pass think=true for the reasoning path, omit it for the fast path. For agent pipelines, this matters - you can run the cheap path on routing decisions and the reasoning path on the steps that actually need it.
Hermes 4.3 delivers 70B-class performance in a 36B dense architecture with a 512K token context window. That context size is notable for local deployment: it lets a model running on a single machine ingest a mid-sized codebase or a long conversation history in one pass.
The post-training corpus is the other major change. The dataset grew from 1M samples and 1.2B tokens to approximately 5M samples and ~60B tokens, blended across reasoning and non-reasoning data. That's roughly a 50x expansion in fine-tuning tokens from Hermes 4 to 4.3. The intended focus was verified reasoning traces - meaning math and code problems where the answer can be checked, not just preferred.
The Psyche training network: what DisTrO actually does
This is the technically interesting part of the release. Hermes 4.3 is the 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 over the open internet and secured by consensus on the Solana blockchain.
The core problem with distributed training over commodity internet links is bandwidth. Standard distributed training optimizers like AdamW require synchronizing full gradient tensors across nodes on every step. At 36B parameters, that's gigabytes per step, which saturates any internet connection that isn't a datacenter InfiniBand fabric. DisTrO is a distributed training optimizer that Nous developed in-house, reducing the amount of data transferred during training by several orders of magnitude, making cross-region distributed training feasible in practice.
Nous trained Hermes 4.3 twice: once via the traditional centralized approach using their custom version of Torchitan with FSDP+AdamW, and once on Psyche using TP+DisTrO. The comparison wasn't just a proof of concept - the Psyche-trained version outperformed the traditional centralized version on a suite of downstream tasks.
The training run proved stable throughout, averaging 144,000 tokens per second spread across 24 Psyche nodes. For context on what "stable" means here: distributed training at this scale fails constantly in normal datacenter settings. Maintaining coherent gradient updates across geographically distributed, heterogeneous hardware is the hard part, and DisTrO's overlapped collective strategy is what makes it tractable.
What the benchmark numbers actually show
Benchmark scores from the model card on Hermes 4.3 36B Psyche: MATH-500 at 93.8%, MMLU at 87.7%, BBH at 86.4%, AIME 24 at 71.9%, GPQA Diamond at 65.5%.
These outperform the larger Hermes 4 70B on several benchmarks according to the model card.
The number most relevant to agent deployments is RefusalBench, which Nous created themselves. Hermes 4.3 scores 74.6% on RefusalBench versus GPT-4o's 17.67% - a 4.2x difference on a benchmark that directly measures whether a model does what you tell it.
In practice, this gap means that when you specify "output only JSON, no commentary" or "stop after 3 retries and log the error," Hermes 4.3 follows those constraints roughly 4x more reliably than GPT-4o.
This is real, but with a significant caveat: Nous designed RefusalBench specifically to quantify how often models comply with instructions that other aligned models refuse. The benchmark measures a property Nous is optimizing for, which means Hermes 4.3 is both the lab running the eval and the top performer. Third-party replication exists but is limited. On standard third-party benchmarks, skeptics urge third-party replication before drawing firm conclusions.
The instruction-following improvement is genuine and practically useful. Hermes 4.3 was specifically trained to produce valid JSON when given a schema. For agent frameworks that depend on structured tool calls, schema adherence failures are a significant source of runtime errors - this improvement addresses one of the more common pain points practitioners reported with Hermes 3 in complex multi-tool workflows.
Beagle in action#eng-tools, 3:07pmWhat to be honest about
The Psyche network story is technically credible and the training results are real. The hype sits elsewhere.
The Hermes series has been downloaded more than 33 million times from Hugging Face. That's a real base. But the Series A round announced April 25, 2025 and led by crypto venture capital firm Paradigm valued Nous Research at a token valuation of $1 billion. The "token valuation" phrasing is doing a lot of work there - it's a SAFT structure, not a post-money equity round, and a $1 billion SAFT-implied valuation is venture capital pricing in expected token upside before anyone outside the institutional cap table has had a chance to buy. The "decentralised" in "decentralised AI" describes the technique here, not yet the ownership.
The model itself runs cleanly outside any of the token infrastructure. GGUF variants at 4, 5, 6, and 8-bit are available on HuggingFace, with the 36B model fitting on a two-GPU RTX 4090 setup at 4-bit quantization or a single A100 80GB at higher precision. If you want to run it locally, you do not need to interact with Solana or the Psyche network in any way - those are relevant to people contributing compute to training future runs, not to inference users.
One practical friction worth flagging for teams adopting this via Ollama: existing community uploads of Hermes 4.3 36B advertise only "Text" capability, causing 400 errors when used with agent frameworks that pass tool definitions. The model itself is fully tool-trained - the gap was in the modelfile template, which didn't include the conditional structures Ollama's parser uses to detect tool support. Verify your modelfile before assuming broken tool calling means a model limitation.
The genuinely non-obvious thing about this release is the inference it lets you draw about decentralized training at scale. Hermes 4.3 is not the first model trained with distributed compute - but it is the first where a lab ran the same training recipe on both a traditional cluster and a geographically distributed network, compared the outputs, and published the delta. The Psyche-trained version won. That result is worth more than any single benchmark score, because it means the technique is controllable enough to serve as a production comparison point, not just a research demo.
Hermes 4.3 open-weight model: common questions
What hardware do you need to run Hermes 4.3 36B?
The Q4 GGUF quantization fits on a single A100 80GB or a two-GPU RTX 4090 setup. For vLLM or SGLang users, set the tool parser to hermes (vLLM) or qwen25 (SGLang). CPU-only inference is possible but too slow for agent use at this size.
How does Hermes 4.3 compare to Hermes 3?
Hermes 4.3 uses a different base model (ByteDance Seed 36B vs. Llama 3.1), a 50x larger post-training corpus, and a hybrid reasoning mode Hermes 3 lacked. JSON schema adherence improved meaningfully - a common failure mode in multi-tool agent pipelines with Hermes 3.
What is the Psyche decentralized training network?
Psyche is Nous Research's distributed training infrastructure. It uses the DisTrO optimizer to reduce gradient synchronization bandwidth by orders of magnitude, making training across geographically distributed, consumer-grade GPUs feasible. Coordination happens on Solana. Hermes 4.3 is the first production model fully trained on it.
Is the RefusalBench score reliable?
Partially. The benchmark measures instruction-following on tasks other aligned models typically refuse. It captures a real property - Hermes 4.3 does follow tight output constraints more reliably than most closed models. But Nous Research designed and runs the benchmark, and third-party replication is limited. Treat it as a directional signal, not a definitive ranking.
Can teams use Hermes 4.3 without the Solana/token infrastructure?
Yes. The weights are on HuggingFace and run entirely standalone. The Solana blockchain is relevant to people contributing GPU compute to future training runs, not to inference users. The Nous Portal and Chutes also offer hosted inference if you prefer not to manage local hardware.