Most model releases don't matter to a business owner. A new checkpoint drops, a benchmark chart circulates, and unless you fine-tune models for a living, it's noise. Hermes 4 is different, not because of hype, but because it's the model running underneath Hermes Agent, the open-source autonomous agent Nous Research built and the one we install and configure for clients. If you're evaluating a self-hosted agent, or you already have one running and want to know what you're paying for under the hood, the details of Hermes 4 stop being trivia and start being a deployment decision. For the wider picture of how it fits into the rest of the ecosystem, see Hermes AI: Why Search Interest Just Spiked. This post stays narrow: what changed, and what it means for the machine you'll run it on.
Three sizes, one model line
Hermes 4 ships in three parameter counts: 14B, 70B, and 405B. All three are built on Llama 3.1 checkpoints, and Nous Research got them to frontier-level performance entirely through post-training, not by training a new base model from scratch. Rather than spending the enormous compute budget a from-scratch frontier run demands, the work went into refining an existing strong base, part of why an open lab can compete with labs that have far more compute to throw around.
The practical read: three sizes means three deployment profiles, not three products. The 14B model is small enough to run comfortably on modest hardware, quick to respond, cheap to host, suited to high-volume, lower-complexity work: triage, drafting, routine lookups. The 70B model sits in the middle, noticeably sharper reasoning, still reasonable to self-host on a serious but not extravagant machine. The 405B model is the heavyweight, the one posting the benchmark numbers below, and it needs hardware to match. None is a "better" version of the others the way a product tier implies. They're sized for different jobs, and picking the right one is closer to picking a server than picking a plan.
Hybrid reasoning, explained without the jargon
The headline feature in Hermes 4 is hybrid reasoning, worth understanding because it changes how the agent behaves, not just how it scores on a leaderboard.
Most models give you one mode: you ask, it answers, it doesn't show its work. Hermes 4 can operate that way too, with fast, standard responses for anything straightforward. But it can also switch into an explicit step-by-step reasoning mode, wrapped in `<think>...</think>` tags, where the model works through a problem deliberately before committing to an answer. The switch isn't guesswork on the model's part: you, or your agent's configuration, control it per task.
Why this matters for an agent: a single session stacks dozens of different tasks together, not one uniform kind of work. A quick reply to a routine email doesn't need deep deliberation, and burning extra tokens and seconds on it is waste. Planning a multi-step research task, working through a pricing model with several dependent variables, or untangling a genuinely ambiguous instruction benefits from the model slowing down first. Hybrid reasoning means your agent isn't stuck at one speed for everything it does. Fast mode handles the volume, reasoning mode gets invoked for the moments that need it, and you decide where that line sits.
How it was actually trained
Nous Research trained Hermes 4 using Atropos, their own open-source reinforcement learning environment. In plain terms: rather than imitating example answers, the model generates attempts, gets scored against roughly 1,000 different task-specific verifiers, and keeps the reasoning trajectories that hold up, discarding the weak ones, across a wide spread of domains rather than one narrow task. What that buys you as the end user is a model pressure-tested against a large, varied set of "did this actually work" checks, part of why Hermes 4 holds up across math, coding, and general reasoning rather than being strong in one lane and hollow everywhere else.
Worth naming too: Nous built Hermes 4 to be creative and genuinely interesting to talk to, describing it as "unencumbered by censorship, and neutrally aligned." A lot of commercial models are tuned toward caution in ways that make them frustrating for legitimate business tasks: they hedge, refuse, or wrap every answer in disclaimers. Hermes 4 is tuned differently while still holding state-of-the-art open-weight performance on the hard technical benchmarks.
The benchmark numbers, and what they mean in practice
Numbers first, since they're the part people ask about. Running in reasoning mode, the 405B model scores 96.3% on MATH-500 and 81.9% on AIME'24, a genuinely hard math competition benchmark that trips up models that look strong on easier tests. Scoring that high in reasoning mode is a real signal, not a rounding trick.
The context that matters more than the raw scores: these results are described as rivaling or exceeding many proprietary systems that cost far more to develop. A closed frontier model is a training run costing a well-funded lab tens of millions of dollars. Hermes 4 is built through post-training on an open base, published with open weights, and lands in the same performance neighborhood on hard reasoning benchmarks. For a business owner, that translates to something concrete: you're not settling for a weaker model because you chose the open, self-hostable, privacy-respecting path.
Worth being precise, since it's easy to overstate: these are two specific benchmarks, in reasoning mode, on the largest model. They tell you Hermes 4 is genuinely capable at hard math and reasoning tasks, not everything about how it'll perform on your specific workflow, which is exactly why the deployment size question below matters as much as the benchmark chart.
Choosing a size is choosing a server
This is where the model release stops being an interesting read and starts being a decision you have to make. Hermes Agent needs a model to think with, and picking which Hermes 4 size it runs on is inseparable from picking the hardware or cloud instance underneath it.
Go with 14B and you're looking at modest, affordable hardware or a small cloud instance, fast responses, and a genuinely capable model for the bulk of day-to-day agent work: triage, drafting, structured lookups, routine tool calls. Go with 70B and you need more capable hardware, but you get sharper reasoning without jumping to the top tier. Go with 405B and you're provisioning serious infrastructure, but you're running the model posting the benchmark numbers above, appropriate if the job involves hard reasoning or technical work where the ceiling matters more than running cost.
This is exactly the same fork we walk through in Hermes Cloud vs On-Device: Privacy & Cost: model size and hosting path are linked decisions, not separate ones. A 405B model on-device means a genuinely serious machine. The same model through a cloud path with scale-to-zero billing changes the calculus considerably, since you're not paying for idle capacity at 3am. Neither answer is right in the abstract; it depends on what the agent needs to do, how often, and how sensitive the data is.
Where this fits into an actual deployment
This matters because it determines what gets provisioned, configured, and secured when a Hermes Agent instance goes live. Picking a Hermes 4 size is one of the concrete decisions we work through with clients as part of our Hermes Agent Installation service, alongside the cloud-versus-on-device call and the tool integrations and access controls the agent needs. The benchmark chart alone can't make this call, and the choice isn't permanent: a 14B deployment that's outgrown its job can move up to 70B without starting over, and the reverse holds if a 405B deployment turns out to be more machine than the workload ever needed.
Hermes 4 is the engine choice for infrastructure you might run, not just an interesting model release. Understanding what changed gives you a real basis for the sizing conversation instead of a guess.
Frequently asked questions
Do I need the 405B model to get good results from Hermes Agent?
No. Most day-to-day agent work, triage, drafting, structured lookups, runs well on the 14B or 70B models, and both cost considerably less to host. The 405B model earns its place when the job involves hard reasoning or technical problem-solving where the extra capability shows up in real output, not just a benchmark chart.
What does "hybrid reasoning" actually change about how the agent behaves?
It gives the model two speeds instead of one. Standard mode answers quickly for straightforward tasks. Reasoning mode, marked by `<think>...</think>` tags, has the model work through a problem step by step before answering, worth the extra time on planning or anything genuinely ambiguous. You, or your agent's configuration, decide when the slower mode gets used.
Is Hermes 4 the same thing as Hermes Agent?
No, they're paired but distinct. Hermes 4 is the model, the part doing the actual reasoning. Hermes Agent is the autonomous system built on top of it, holding memory, connecting to tools, and doing persistent work across a session. The agent needs a model to think with, and Hermes 4 is that model, in whichever size fits your deployment.
"Rivaling proprietary systems" sounds like marketing. Is it?
The specific numbers behind it are real and narrow: 96.3% on MATH-500 and 81.9% on AIME'24 for the 405B model in reasoning mode, benchmarks that are genuinely hard to score well on. That's not the same as claiming Hermes 4 beats every closed model at everything, and we'd rather you go in with accurate expectations than an inflated one. If you want help working out which Hermes 4 size fits your workload and hardware budget before committing, that's exactly what the deployment assessment inside our Hermes Agent Installation service covers. Write to team@ins.ae or message us on WhatsApp at +971 58 995 4553 and we'll walk through it with you.
