You can run Hermes Agent — Nous Research's MIT-licensed, self-improving AI agent — for zero ongoing cost by pointing it at Tencent's Hunyuan 3 (Hy3) model through a free OpenAI-compatible endpoint. The two pieces are independent: Hermes is the agent runtime that calls tools and remembers context; Hy3 is the 295-billion-parameter reasoning model that does the thinking. As of July 2026, Novita AI serves the official tencent/hy3 model at $0 per million tokens (input and output) on a serverless, OpenAI-compatible API, so the only requirement is a small onboarding credit to unlock access. This guide wires those two together end to end and then gives you a job framework so your first agent actually finishes instead of stalling.
Last verified: 2026-07-24 · Free route: Novita AI (
tencent/hy3, $0/M launch pricing) · Model: Tencent Hy3, 295B MoE / 21B active / 256K context · Runtime: Hermes Agent (MIT, Nous Research). Pricing and free-tier availability are volatile — re-check the linked provider pages before relying on this for production.
What is Hermes Agent and why does the model choice matter?
Hermes Agent is an open-source AI agent framework from Nous Research (GitHub: NousResearch/hermes-agent, MIT license). Unlike a chatbot, it has a built-in learning loop that creates reusable skills from experience, a three-layer memory system so context survives across sessions, a cron scheduler for recurring jobs, and support for more than 20 messaging surfaces (Telegram, Discord, Slack, CLI, desktop, email). Crucially for this setup, it works with any OpenAI-compatible endpoint — you point it at a base URL, give it a key, and it treats that as its model (Hermes providers docs).
The model behind the agent is what decides whether a 20-minute job finishes or stalls. Agents burn through context fast (a job that reads three documents and makes ten tool calls eats more memory than a week of normal chat) and they depend on reliable tool calling — picking the right tool, sending the right arguments, and remembering the original instruction after the result comes back. Hy3 was tuned specifically for those two failure modes. Tencent's official release notes describe "enhanced agent capabilities" with improved "multi-turn dialogue, and agentic task execution" (Tencent corporate news, July 6, 2026), and the preview scored 74.4 on SWE-bench Verified and 54.4 on Terminal-Bench 2.0 — both agentic coding benchmarks (Novita AI model page).
How to get Hunyuan 3 (Hy3) for free: the 3 routes compared
There are three free or near-free ways to reach Hy3, and which one you pick depends on whether you want zero setup, scriptable API access, or full data control. All three speak the OpenAI chat-completions protocol, so switching between them is a one-line change in Hermes.
| Route | Provider / host | Cost as of Jul 2026 | What you get | Best for |
|---|---|---|---|---|
| 1. Novita AI (free API) | Novita AI (nova.ai) |
$0/M tokens launch rate; needs ≥$1 balance to unlock | Serverless OpenAI-compatible API, 262K context, function calling + structured outputs + 3 reasoning modes | The recommended free path — see setup below |
| 2. OpenRouter (paid) | OpenRouter | tencent/hy3 now ~$0.14/M input, $0.58/M output (the free :free variant was deprecated 2026-07-21) |
Same model, OpenAI-compatible, one billing account for 100+ models | Builders already routing through OpenRouter |
| 3. SiliconFlow (free 2 weeks) | SiliconFlow | Free for the first 2 weeks from Jul 5, 2026, then paid | OpenAI-compatible API, hosted Hy3 | Trying Hy3 on a second provider for comparison |
| 4. Self-host (vLLM/SGLang) | Your own GPUs | Free software; ~8× H100-class GPUs for the full BF16 model | Full control, no rate limits, no data leaves your box | Privacy-sensitive or high-volume |
The cleanest free path today is Novita AI. Novita's model page confirms tencent/hy3 is served at $0/M input and $0/M output with a 262,144-token context window and full function-calling support (Novita AI Hy3 model page); the Novita blog post confirms this is the launch pricing (Novita blog: Hy3 on Novita AI). The one catch, stated by Novita directly: the model is free but you need a small positive API balance to access it, and completing onboarding earns a $1 API credit which unlocks the free endpoint (Novita AI on X/Twitter). One dollar of balance is plenty for unlimited calls to a $0/token model — the balance is the gate, not the meter.
How to set up Hermes Agent with Hunyuan 3 on Novita (step by step)
This is the shortest path from nothing to a working agent call. Both pieces are free; total setup is under 15 minutes if you have a developer machine.
Step 1 — Install Hermes Agent
Hermes ships a one-line installer for macOS, Linux, and Windows (Hermes Agent homepage):
# macOS / Linux
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
# Windows (PowerShell)
irm https://hermes-agent.nousresearch.com/install.ps1 | iex
A desktop app is also available from the same page if you prefer a GUI. Hermes is MIT-licensed, so there is no paid tier to unlock.
Step 2 — Get your Novita account and a $1 onboarding credit
- Sign up at novita.ai (GitHub or Google login works).
- Complete the onboarding checklist in the dashboard. Novita confirms this grants a $1 API credit that unlocks the free Hy3 endpoint (Novita on X).
- Generate an API key in Dashboard → API Keys. Copy it; you'll use it in the next step.
Step 3 — Point Hermes at Novita's Hy3 endpoint
Hermes treats Novita as a custom OpenAI-compatible endpoint. Run the model wizard and enter these exact values (verified against the Novita Hy3 model page and the Hermes providers docs):
hermes model
# Select: Custom endpoint (self-hosted / VLLM / etc.)
# Base URL: https://api.novita.ai/openai
# API key: <your Novita key>
# Model name: tencent/hy3
Or, if you prefer editing config directly, the equivalent in ~/.hermes/config.yaml is:
model:
default: tencent/hy3
provider: custom
base_url: https://api.novita.ai/openai
Hermes verifies the endpoint by probing /models on save, so if you mistyped the URL it will tell you immediately rather than failing silently at runtime (the Hermes GitHub issue #1460 confirms this verification step ships in the current flow).
Step 4 — Send your first test message
hermes chat
# Inside the chat:
> What's 17 × 23, and break down how you got it.
If you get a coherent step-by-step answer, Hy3 and Hermes are talking. If you get a 429 (rate limited), wait a moment — Novita's free pool is shared and peak-hour calls can be throttled (Novita Hy3 blog).
Step 5 — Pick the right reasoning mode for the job
Hy3 exposes three reasoning modes via the reasoning_effort parameter (Novita Hy3 model page):
| Mode | When to use it | Trade-off |
|---|---|---|
no_think |
Quick routing, simple completions, chat | Fastest, no reasoning tokens |
low |
Light multi-step reasoning | Slightly slower, better quality |
high |
Long agentic chains, code, math | Slowest, best quality; reasoning tokens are billable on paid routes |
For your first job, leave it on the default and only bump to high when you notice the agent struggling on a multi-step task. Reasoning tokens are free on the Novita $0 launch pricing, but if you later switch to a paid route (OpenRouter, Tencent Cloud) they bill as output tokens — so don't ship high as a default in production code without measuring.
The 3-part "job, tools, memory" framework for writing agents that finish
A working agent is not just a model and a runtime — it's a written job, a constrained toolset, and persistent memory. Three pieces, in that order. Get all three right and your agent finishes tasks; skip any one and it stalls or hallucinates halfway through.
1. Write the job like an onboarding doc, not a wish
The single most common reason an agent fails is a vague instruction. "Clean up my content calendar" tells the agent nothing — it will guess what "clean" means and guess wrong. Write the job as if you're handing it to a smart contractor on their first day: which files to open, what to look for, what the output should look like, and where to save it.
Weak job: "Summarize our community questions."
Strong job (rewrite, concrete):
Read
/inbox/questions.json. Group all questions by topic. Count how many people asked about each topic. List the top five. For each of those five, write one sentence describing what people are actually confused about. Save the result toweekly-topics.txtand print the path.
Same intent, completely different result. The agent now has a deterministic shape to fill in instead of a vague goal to interpret.
2. Start with three tools, not fifteen
The second classic mistake is giving an agent every tool you have on day one. An agent with fifteen tools and a fuzzy job picks tools at random and burns twenty minutes doing nothing useful. Start narrow — three tools — get the job working, then add one tool at a time as the job proves it needs them. If you want a deeper treatment of this deprecation-style approach, our how to organize your AI work like an operating system guide walks through the tool-selection layer in detail.
3. Build in a checkpoint before the final output
The third mistake is the most expensive: letting the agent run for twenty minutes and hand you a result that went wrong at minute two. Build in a stop. Have the agent print its plan before it starts, or write out a draft of what it found before it writes the final output. Ten seconds of your attention catches a problem that would otherwise cost ten minutes of rework.
A practical example: an agent that turns a long transcript into a tutorial should first print a numbered list of the steps it found. You glance at the list, approve or correct, and then it writes the full guide. This same checkpoint pattern is what makes multi-agent Hermes Agent workflows practical — the supervisor reviews a sub-agent's plan before letting it execute.
The 20-minute test: how to know your free setup actually works
Short tasks make every model look fine. The real test of an agent + model pairing is a 20-minute job — something long enough that the model has to hold the original instruction across many tool calls, recover from at least one failed step, and produce a coherent final artifact. Run one this week, while Hy3 is free, and you'll learn more about whether this setup fits your business than any benchmark number will tell you.
A good 20-minute job for your first run: have the agent read every piece of content you've published in the last month, identify the three most common themes, and write a one-paragraph "actual voice" description based on the words you really use (not what you think your voice is). This job needs long context (it has to hold a lot at once) and it needs tool calling (it has to read files and write a result) — exactly what Hy3 was tuned for. It's also the kind of job people normally hesitate to run because the meter is running. With a $0/M model, you just run it.
If the run succeeds — the agent remembered the original instruction after multiple tool calls, stayed on task, and produced something usable — you have a trustworthy pairing. If it stalled, check the model (try reasoning_effort: high) before you blame the framework. For a fuller framework on what to actually delegate to agents and how to scale that into a real business operation, see our how to run a small business on AI agents guide.
How to switch providers when the free tier ends
Free launch pricing does not last forever. Novita's $0/M rate is a launch promotion and could change; OpenRouter's free tencent/hy3:free endpoint was already deprecated on 2026-07-21 (the paid tencent/hy3 remains, at ~$0.14/M input and ~$0.58/M output per APIRank's coverage); SiliconFlow's two-week free window started July 5, 2026. The good news: because every route speaks the OpenAI-compatible API, switching is a one-line base_url + model change in Hermes. You never rewrite your agent jobs.
When the free window closes, your options are:
- Stay on Novita paid — keep the same
base_url, the samemodelID; you just start paying per token once the $1 credit and the $0 launch rate end. - Move to Tencent Cloud directly — the official Hy3 is on Tencent Cloud TokenHub at roughly ¥1 per million input tokens (~$0.14/M) per Tencent's July 6 release; the Tencent Cloud Hunyuan pricing page has the exact rate card.
- Self-host the open weights — Hy3 is open-weight under the Tencent Hy Community License (Hugging Face: tencent/Hy3-preview). Self-hosting is free software but needs roughly 8× H100-class GPUs for the full BF16 model per the official vLLM quickstart; quantized community builds exist for smaller setups.
- Local and offline, no cloud at all — if you want true zero-cost with no API dependency, how to run Gemma 4 locally as a free coding assistant is the same Hermes-custom-endpoint pattern pointed at an Ollama base URL (
http://localhost:11434/v1) instead of a cloud host.
What this means for you
If you've been hesitant to run long agent jobs because every step has a meter running, this week is the testing window. Plug Hermes Agent into the free Hunyuan 3 endpoint on Novita, write one concrete 20-minute job (job + three tools + a checkpoint), and run it. You'll learn in one evening whether agent automation fits your work — for zero cost — instead of theorizing about it. Bookmark this page and re-check the Novita and OpenRouter pricing pages before you rely on it for anything customer-facing; free launch rates are volatile and the OpenRouter free variant has already ended.
FAQ
Q: Is Hermes Agent actually free?
A: Yes. Hermes Agent is open-source under the MIT license and installable with a one-line curl command. There is no paid tier required to run it — the cost comes from the model API you point it at, which is why pairing it with a free endpoint like Novita's tencent/hy3 matters.
Q: Is the Hunyuan 3 (Hy3) free endpoint permanent?
A: No. Novita's $0/M rate is explicitly a launch promotion as of July 2026, and the OpenRouter free variant (tencent/hy3:free) was already deprecated on 2026-07-21. Treat the free window as a testing opportunity, not a production plan. The paid tencent/hy3 endpoint remains on OpenRouter at ~$0.14/M input and ~$0.58/M output, and Tencent Cloud serves the official model at ~¥1/M input tokens.
Q: Do I need a credit card to use the Novita free Hy3 route? A: You need an account and a small positive API balance. Novita states that completing onboarding grants a $1 API credit which unlocks the free Hy3 endpoint — the balance is the gate, not a per-call meter, since the model itself is $0/M tokens (Novita on X).
Q: Why use Hunyuan 3 instead of a free local model like Gemma 4? A: Context and tool-calling reliability. Hy3 has a 256K-token context window and was specifically tuned for agentic tool use and long multi-step jobs — exactly where small local models tend to stall. A free local model is the right choice when you need offline privacy or zero API dependency; Hy3 is the right choice when you need a big-context, tool-reliable model and don't mind a cloud round-trip. See our Gemma 4 local setup guide for the offline alternative.
Q: Can I run Hy3 locally instead of through an API? A: Yes — the model weights are open under the Tencent Hy Community License on Hugging Face. The official quickstart uses vLLM or SGLang with 8× H100-class GPUs for the full BF16 model; quantized community builds can run on less but trade quality.
Q: What's the difference between Hermes Agent and a chatbot? A: A chatbot takes a prompt and returns text. An agent takes a job, picks up tools (reads a file, opens a browser, sends an email), uses memory to carry context across runs, and finishes a task instead of answering a question. Hermes Agent is built around that job + tools + memory model — see our how to build an AI agent team with Hermes Agent guide for the full architecture.
Discussion
0 comments