Verdict: Tencent Hy3 is a 295-billion-parameter open agent model that ran completely free inside Hermes Agent (via Nous Portal) and on OpenRouter for roughly two weeks in July 2026. That promotional window has closed, but Hy3 remains a strong open-weight option for agent work, and several genuinely free agent-capable models still exist on OpenRouter today. This guide covers what Hy3 actually is, what the free period meant, what is still free, and how to plug any of them into Hermes Agent in a few minutes.
TL;DR
- Tencent Hy3: 295B total / 21B active MoE, 256K context, Apache 2.0 open weights
- Free Nous Portal + OpenRouter window ran early-to-late July 2026 — it is now closed
- Hermes Agent setup takes one command (
hermes model) plus picking the model- Still free on OpenRouter right now: KAT-Coder-Air V2.5, LFM2.5-2.6B, and ~23 others
- Last verified: 2026-08-24 — pricing and free-tier lists change often
What exactly is Tencent Hy3?
Tencent Hy3 is a 295-billion-parameter Mixture-of-Experts (MoE) model with 21 billion active parameters per token (top-8 of 192 experts), a 256K token context window, and Apache 2.0 open weights, released by Tencent in early July 2026 (Tencent Hy3 GitHub repository). It is the production successor to Hy3 Preview from April 2026 and is explicitly tuned for agent workloads: coding, tool calling, reasoning, and long-context tasks.
A few numbers from Tencent's own published results:
| Benchmark | Hy3 score |
|---|---|
| SWE-bench Verified | 78.0 |
| SWE-bench Pro | 57.9 |
| GPQA Diamond | 90.4 |
| HLE | 53.2 |
(Vendor-reported figures; treat as directional, July 2026.)
The interesting architectural detail for agent builders is the cost pattern: only 21B parameters fire per token, so each step of a long agent loop costs roughly what a 21B dense model would, even though the model has 295B total capacity. That is why Nous Research positioned it as a cost-efficient brain for Hermes Agent rather than a chatbot.
Did Hy3 really run free inside Hermes Agent?
Yes — for about two weeks starting July 6, 2026, Nous Research made Hy3 free inside Nous Portal (the subscription gateway for Hermes Agent), and OpenRouter listed it as :free until approximately July 21, 2026 (OpenRouter free-model listing via a community tracker, July 2026). Both windows have now closed. If you saw headlines about a free 295B agent model free for 2 weeks — that was this one, and it has ended.
Hermes Agent itself is always free — it is Nous Research's open-source (MIT) autonomous CLI agent with persistent memory, self-improving skills, and a messaging gateway covering 21+ platforms including Telegram, Discord, and Slack (Hermes Agent docs). What ended was the free model, not the free agent.
What free agent models can you still use?
As of August 2026, OpenRouter's free collection still holds about 25 models; the most agent-relevant entries are (OpenRouter free models):
| Model | Context | Notable for |
|---|---|---|
| KAT-Coder-Air V2.5 | 256K | Agentic coding, tool calling — strongest free coding option |
| Qwen3 Coder 480B-A35B | 1M | Long-context agentic coding |
| DeepSeek V4 Flash | 1.3M | Extremely cheap reasoning for agent loops |
| NVIDIA Nemotron 3 Nano Omni (30B) | 256K | Multimodal perception sub-agent |
| Liquid LFM2.5-2.6B | 128K | Tiny edge / always-on agent tasks |
Free-tier listings rotate with little warning — always check OpenRouter's free-collection page before building on one. Our own guide to free DeepSeek V4 Flash access and the broader best open-source LLM for coding comparison cover the current alternatives in more depth.
How do you connect one of these to Hermes Agent?
- Install Hermes Agent if you have not already — see our Hermes Agent setup guide for the full install path.
- Pick your provider: OpenRouter (bring-your-own key, free models included) or Nous Portal (one subscription across the whole model catalog).
-
Run the model picker: say
hermes modelin your terminal, pick the provider, then pick the model (for exampletencent/Hy3or a free:freeentry). That is all Hermes needs — no YAML edits required for standard setups. -
If you use OpenRouter directly, get a free key from openrouter.ai, add a credit minimum only if you plan to use paid models, and pass the model ID (for example
tencent/hy3orqwen/qwen3-coder:free) through the picker. OpenRouter's own integration cookbook documents this flow (OpenRouter + Hermes Agent).
If you want a model that does not appear in the picker, point Hermes at the OpenRouter base URL manually; the only hard rule is that the model needs at least 64K context for the agent harness to start cleanly.
What can a free-ish agent model actually do inside Hermes?
More than most people expect from a non-frontier model — with the right task mix:
- Tool use in real time. Open a URL, read a page, write a file. Hy3-class models handle the standard tool-call loop on OpenRouter without extra setup.
-
Learn mode. Type
/learn, paste a link to a guide, and Hermes distills it into a reusable skill it keeps for future sessions. This compounds: the agent gets better at your business over time even on a cheaper brain. Our article on self-improving agents in Hermes covers the pattern in depth. - Background tasks. Give it a job, walk away — the agent keeps running. On a slow free model this is the right pattern; on a fast paid model you would use lighter prompting instead.
The trick is mixing tiers: run a free model for routine, low-stakes steps (summaries, drafts, small file edits) and keep a frontier model for the hard reasoning at the top of the loop. Our OmniRoute free gateway guide shows one way to wire several free models behind a single agent so rate limits on one automatically fall back to another.
What this means for you
If you have been waiting for a free 295B agent brain, you missed the Hy3 window — but the gap is smaller than it looks. Free coding-agent models like KAT-Coder-Air V2.5 are still good enough to power real workflows inside Hermes Agent, and DeepSeek V4 Flash at $0.07/M input is close enough to free for most small-business use. Set up Hermes once, point it at OpenRouter, and you can swap brains whenever a new free promotion drops — Nous runs them regularly.
FAQ
Q: Is Tencent Hy3 still free anywhere in 2026?
A: No — the two-week Nous Portal and OpenRouter :free promotions ended in late July 2026. Hy3 is now a paid model on OpenRouter, though still very cheap relative to frontier models because of its 21B-active MoE design.
Q: Can I self-host Tencent Hy3 instead of paying for the API?
A: Yes. Weights are on Hugging Face under Apache 2.0, and both vLLM and SGLang have dedicated recipes. You need about 8 high-memory GPUs, so this is only economical at scale.
Q: What is the best free model for Hermes Agent right now?
A: For pure coding and tool use, KAT-Coder-Air V2.5 is the strongest fully-free option on OpenRouter. For general reasoning loops, DeepSeek V4 Flash is free on some routes and nearly free on the rest. Check OpenRouter's free models page — the lineup rotates.
Q: Do I need a Nous Portal subscription to use Hermes Agent?
A: No. Hermes Agent is open source and free to run; you can point it at OpenRouter, direct API keys, local models, or Portal. Portal just bundles many providers behind one login.
Q: What happens when a free model gets rate-limited mid-task?
A: Switch profiles. Hermes can run one profile on model A and another on model B; if the first gets throttled you flip over and continue. Genuinely free agent gateways like OmniRoute automate this fallback.
Sources
- Tencent Hy3 on GitHub — model card, benchmarks, license
- Tencent Hy3 on Hugging Face — open weights, variants
- Mervin Praison: "Tencent Hy3: 295B MoE Agent Model Free on Nous Portal" (July 7, 2026) — recap with Nous announcement link
- Pinggy: free AI model APIs on OpenRouter — Hy3 free-window end date (July 2026)
- OpenRouter: free models collection
- OpenRouter: Hermes Agent integration cookbook
- Hermes Agent — official docs
Last verified: 2026-08-24 · Updates: 2026-08-24 — published; recorded Hy3's July free window as closed and current OpenRouter free-model alternatives.
Researched and drafted with AI agents; fact-checked against the primary sources above under human editorial oversight. See how we work for our methodology.
Top comments (0)