Hermes Agent Step 5 means running StepFun's new open-weights model, Step 5 Preview, as the brain inside Hermes Agent. It's built for agent work, it's fast, and when I recorded my video on 24 Oct 2026 it was free to run inside Hermes. I used it to learn a new skill, run research, draft content and hold a spoken conversation, all without touching my premium model budget.
📺 Watch: Hermes Agent + Step 5 Is the BEST FREE AI Combo?!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
Here's what the model is, how to plug it in, what I ran on it, and which jobs I still hand to a frontier model.
Want my exact Hermes setups and the full Agent OS? They're inside the AI Profit Boardroom ($69/mo).
What Step 5 Preview is
StepFun calls Step 5 Preview its "flagship model for agentic work." It's a sparse Mixture-of-Experts model. That means it has a huge pool of parameters but only switches on a small slice of them for each token. That's why it feels quick for its size.
The weights are public on Hugging Face under stepfun-ai, in an FP8 build and a BF16 build, plus ModelScope copies. The main repo went up on 20 Sep 2026, and an NVFP4 build followed on 14 Oct 2026.
Here are the specs StepFun has published, from the model card and the StepFun platform docs.
| Spec | Step 5 Preview | Source |
|---|---|---|
| Architecture | Sparse Mixture-of-Experts | Hugging Face model card |
| Total parameters | 600B | Model card |
| Active parameters | 27B per token | Model card |
| Context window | Up to 1M tokens | Model card and StepFun docs |
| Max output on StepFun's API | 64k tokens | StepFun docs |
| Input / output | Text and images in, text out | Model card |
| Reasoning effort | low, medium, high | StepFun docs |
| Tool calling | Supported | StepFun docs |
| API model name | step-5-preview | StepFun docs |
| List price on StepFun | $1.00 per 1M input tokens ($0.10 cached), $2.70 per 1M output | StepFun pricing page |
| Licence | StepFun Model License 1.0 | Model card |
On benchmarks, StepFun's own chart puts Step 5 Preview (on High reasoning) at 67.7 on DeepSWE v1.1, 80.5 on ProgramBench, 33.3 on Terminal-Bench v4 and 29.5 on Agents' Last Exam (ALE-CLI). In that same chart, Claude Opus 5 scores 74.0 on DeepSWE and 52.3 on Terminal-Bench, and GPT-6 Astra scores 74.1 and 57.9. So StepFun's numbers show it close to Kimi K3 and GLM-5.3 on coding, and still behind the top frontier models on the hardest terminal work.
The licence is free to use, copy and modify. StepFun describes it as Apache-2.0-style with extra conditions: you show "Powered by StepFun AI" once a product passes 100 million monthly users or $20 million in monthly revenue, and device makers and big model-hosting businesses need a separate commercial licence. Internal use doesn't trigger those conditions.
If you want the full Hermes setup I use, the Agent OS, the 30-day AI agent roadmap and four coaching calls a week are inside the AI Profit Boardroom ($69/mo). Over 3,400 members are building agents in there right now.
Why pair Hermes Agent with Step 5
Most of what an agent does all day is not hard thinking. It's reading pages, summarising, filing notes, rewriting drafts, calling tools and checking results. You don't need your most expensive model for that.
That's where Step 5 fits. It's lightweight on purpose. It's not trying to be the smartest model on earth. It's trying to be fast and good enough for the everyday loop, so your frontier credits go to the jobs that actually need them.
Hermes makes this easy because I run a profile per model inside my Agent OS. One profile on Step 5 for the daily grind. Another on a frontier model for the heavy lifting. Same skills, same memory, different brain. If you're new to that setup, my Agent OS guide walks through it.
It's the same play I made with small local models like Spark-X2.6-4B, just with a much bigger model doing the cheap work.
📺 Watch: Run Hermes Agent FREE With Step 3.5 Flash 🤯
How to set up Hermes Agent Step 5
Hermes already supports StepFun as a built-in provider. The Hermes docs list it as "stepfun", using an OpenAI-compatible API with key authentication.
- Get a StepFun API key from platform.stepfun.com.
- Add it to your Hermes env file at ~/.hermes/.env as STEPFUN_API_KEY.
- Start a chat with the StepFun provider and the Step 5 model name: hermes chat --provider stepfun --model step-5-preview
- To make it your default, run hermes model and pick StepFun, or set provider "stepfun" and default "step-5-preview" under model in config.yaml.
The Hermes docs show step-3.5-flash as the example model. Swap in step-5-preview, which is the model name in StepFun's API docs. The default base URL is https://api.stepfun.com/v1, and you can override it with STEPFUN_BASE_URL if you need a different region.
Nous Research's own model list also carries stepfun/step-5-preview, so if you're on Nous Portal you can pick it from hermes model there too.
Want this wired into your business without the trial and error? Grab a free AI strategy session and we'll map out which agent jobs to move first.
📺 Watch: Hermes3D Is INSANE: AI Agents Now Have a 3D Office
What I ran on it
I put Step 5 through the kind of work I actually do every day. Four things stood out.
Teaching it a skill
I taught it a new skill in seconds. I explained what I wanted, Hermes wrote the skill, and from then on the agent could reuse it. Hermes handles this with its skill_manage tool, so the skill sticks around for the next session instead of living in one chat.
Research with live web search
I ran research with live search built in. Step 5 pulled sources, read them and gave me a tidy summary. With a 1M-token context window, it can hold a lot of source material in one go.
Drafting
Then I had it draft content from that research. This is exactly the job where a cheaper model earns its keep. Drafts get rewritten anyway, so you want speed and volume, not perfection.
Talking to it out loud
I also talked to it out loud. In Hermes, you set up voice providers with hermes tools, type /voice on, and press Ctrl+B to record. Type /voice tts if you want it to talk back. Step 5 kept up fine for conversation.
Under all of that sat shared memory. I use Obsidian as the memory layer, so every agent in my Agent OS sees the same notes no matter which model it runs on. My Hermes Obsidian page shows that setup. I also chained agents together, so one researches, one drafts and one formats, all on the cheap model.
Every member of the AI Profit Boardroom gets the prompt library and the daily tutorials, so you can copy these workflows as new models drop.
Jobs to keep on a frontier model
Step 5 is a workhorse, not your top gun. StepFun's own chart shows a real gap on hard terminal and coding work. On Terminal-Bench v4 it scores 33.3 against 52.3 for Claude Opus 5. On StepCodeBench it's 49.0 against 63.9.
So here's how I split it:
- Keep on Step 5: research summaries, first drafts, note filing, rewriting, simple tool chains, voice chats, repeat skills.
- Move to a frontier model: long multi-step coding in a real repo, tricky terminal and server work, final client-facing copy, anything where one mistake costs you money.
For the heavy jobs I swap to a stronger profile, like the one in my Kimi K3.1 Pro Hermes setup, or a top closed model. I also test models on my own tasks with Goldie Bench before I trust them with real work.
If you're not sure which of your jobs belong where, book a free AI strategy session and we'll sort it with you.
Limits to know
- The free window is temporary. It was free inside Hermes when I recorded. StepFun's list price is $1.00 per million input tokens and $2.70 per million output, so check your provider before you lean on it at scale.
- It's a preview. The name says it. Expect changes.
- Self-hosting is not a laptop job. The model card's smallest setup is FP8 on one node of four B300 GPUs or eight H200 GPUs. Open weights are great, but most people will use the API.
- Output length differs by route. The model card says up to 1M tokens of output, but StepFun's own API caps output at 64k.
- Licence conditions kick in at scale. Fine for most businesses, but read it if you're building a big product or a hosting service.
FAQ
Is Step 5 free in Hermes Agent?
It was free to run inside Hermes when I recorded on 24 Oct 2026, and that kind of window doesn't last. The weights themselves are free to download under the StepFun Model License 1.0. On StepFun's own API, it costs $1.00 per million input tokens and $2.70 per million output.
What's the model name to use in Hermes?
Use the stepfun provider with step-5-preview as the model. Put STEPFUN_API_KEY in ~/.hermes/.env first.
Can I run Step 5 on my own computer?
Not on a normal one. StepFun's deployment guide starts at a full GPU node, like four B300s or eight H200s for the FP8 build. Use the API unless you run serious hardware.
Should Step 5 replace my frontier model?
No. Use it for the everyday agent loop and keep a frontier model for hard coding and high-stakes work. Two profiles in Hermes make that a one-command swap.
Ready to build your own agent stack around Hermes Agent Step 5? Book a free AI strategy session and we'll set up your first profiles together.
