A 1.6-trillion-parameter frontier model just went free for a week, and it plugs straight into the agent you already run. That is the Hermes LongCat story in one sentence: Meituan's LongCat team released LongCat-2.0, Nous Research put it in their portal at no cost for seven days, and because Hermes swaps brains the way you swap batteries, you can have it doing real work for you today. Here is what the model actually is, why it fits agent work unusually well, and exactly what I would test before the free week ends.
📺 Watch: Owl Alpha: New FREE AI Stealth Model!
🔥 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
The offer first, because it is time-limited. LongCat-2.0 is free in Nous Portal for one week. Nous Research is the company behind the Hermes agent, which is what makes this more than another model launch: the same ecosystem your agent lives in is handing you a frontier-class brain to run through it, free, for seven days. No waitlist drama, no new subscription. Sign in, pick the model, point your agent at it.
What LongCat-2.0 Actually Is
LongCat-2.0 is a mixture-of-experts model with 1.6 trillion total parameters and roughly 48 billion active at any moment. In plain English: the full network is enormous, but for each token only the relevant slice of experts wakes up and does the thinking. The LongCat team push this further with what they call Zero-Compute Experts — activation flexes dynamically between 33 billion and 56 billion parameters per token, so easy tokens spend less compute and hard tokens get more. Their phrase for it is zero wasted compute.
The second headline is context: a 1-million-token window. That is enough to ingest an entire codebase in one pass — not summaries of your repo, the repo. Their attention design, LongCat Sparse Attention, is what lets the model scale to that window without grinding to a halt. And the third piece is the one I find most interesting for agents: MOPD, three specialised expert groups — Agent, Reasoning and Interaction — with a gate that routes each task to the right group. This is a model that was built for agentic coding from the ground up, not a chat model with agent features bolted on afterwards.
The Owl Alpha Reveal
If LongCat-2.0 feels like it came out of nowhere, it did not. It is the full model behind Owl Alpha — the stealth model plenty of people have been quietly using on OpenRouter and rating highly without knowing whose it was. The mask is now off. If you were one of the people who ran Owl Alpha and liked it, you already have hands-on experience with this model; the free week is your chance to run the real thing with a name on it.
If you want new brains like this wired into your agent the week they drop, the AI Profit Boardroom covers swaps exactly like this inside the Agent OS as they land — so a model release becomes a profile change, not a rebuild. → Get the setup
📺 Watch: Hermes AI Agents Just Went Portable
The Numbers LongCat Published
These are the LongCat team's own published figures, so treat them the way I treat every launch benchmark — as the lab's best foot forward until independent testers confirm them. With that said, they are striking. Terminal-Bench 2.1, which measures how well a model works inside a real terminal and finishes real tasks: 70.8. SWE-bench Pro: 59.5 — a nose ahead of the 58.6 they report for GPT-5.5 on the same test. SWE-bench Multilingual: 77.3. On the research-and-browsing side: 73.2 on FORTE, 78.8 on RWSearch and 79.9 on BrowseComp. The pattern across the board is exactly what the architecture promises: strong agentic coding, strong tool-driven research, built to run long jobs rather than win chat-window beauty contests.
One caveat worth repeating: the lab picked the tests. Every lab does. That is not an accusation, it is the reason the free week matters — you get seven days to run your own tests on your own work before the meter exists.
The Shape Of The Agent Era
Step back and there is a pattern here. Days before this release, DeepSeek shipped V4 Pro at almost identical dimensions — 1.6 trillion total parameters, around 49 billion active, a 1-million-token window. Two frontier labs, same month, same shape: sparse giants with huge context, engineered for agents that re-read big context constantly and grind through long multi-step tasks. I broke that model down in my DeepSeek V4 Pro vs Claude Fable 5 vs Grok 4.6 comparison, and LongCat-2.0 slots straight into the same conversation. The age of one dense model for everything is over; the agent era belongs to models like this.
📺 Watch: Grok Bot DESTROYS Hermes Agent?
Why Hermes LongCat Is The Right Pairing
Hermes is provider-agnostic by design. Your agent's identity — the persona, the skills, the memory, the schedules — lives in Hermes, and the model underneath is just a profile setting. Swap the brain, keep the agent. That design is precisely why a release like this is exciting rather than disruptive: nothing about your setup changes except the quality of the thinking.
And the model's shape fits agent work unusually well. Agents re-read their context on every step, which is where a 1M window stops being a spec-sheet flex and starts being the difference between an agent that holds your whole project in mind and one that keeps forgetting the start of it. Feed it the entire codebase in one pass and ask for the refactor. Hand it a season of research notes and ask for the strategy. The MOPD gate routing agent tasks to agent experts is exactly the workload Hermes generates all day. If you have been running a free or budget brain behind Hermes — I keep a list in best free AI model for Hermes agent — this week is a straight upgrade at the same price: nothing.
How To Try It This Week
- Go to portal.nousresearch.com and get access to LongCat-2.0 while the free week is running.
- Point a Hermes profile at it as the model — the same way you would select any brain. If you keep a profile per model like I do, make it a fresh profile so histories stay separate.
- Give it your real work, not demos: the tasks you currently hand your daily driver.
- Judge outputs side by side before the week ends, so the keep-or-drop decision is made on evidence.
What I Would Test Before The Week Ends
This is where my Goldie Bench habit earns its keep: same task, two brains, judge the output — never take a launch chart's word for it. For LongCat-2.0 the three tests that matter are, first, a real coding task from your actual backlog, because Terminal-Bench numbers mean nothing until it fixes your build; second, a genuine long-context job — drop a whole project or a giant document set in one pass and probe whether it can still find the details buried in the middle; third, an agent loop with tool calls, because a model built for agentic work should shine precisely where chat models fall apart — multi-step tasks with tools, memory and recovery from its own mistakes. Run those three against whatever brain you use today and the free week will have paid for itself in information.
After The Free Week
When the week ends, the model does not vanish. It has been running on OpenRouter as Owl Alpha, so a routed API path exists beyond the portal — and if it earns a place in your rotation, that is where it slots into the wider free-and-cheap stack I covered in free API for Hermes agent. The usual cloud-model note applies as ever: hosted frontier models run on someone else's servers, so keep genuinely sensitive work on the setups you control — my Hermes local model setup guide covers that side.
| Spec | LongCat-2.0 |
|---|---|
| Total parameters | 1.6 trillion, mixture-of-experts |
| Active per token | ~48B, flexing 33B–56B via Zero-Compute Experts |
| Context window | 1 million tokens — a whole codebase in one pass |
| Built for | Agentic coding, with Agent / Reasoning / Interaction expert groups |
| Headline scores (lab-published) | 70.8 Terminal-Bench 2.1 · 59.5 SWE-bench Pro · 79.9 BrowseComp |
| Where it is free | Nous Portal, for one week |
For the wider context on where this model line came from, my earlier LongCat-2.0 vs GLM-5.2 comparison covers how it stacks against China's other open flagship, and the best open source models for Hermes agent ranking shows the field it is entering.
Hermes LongCat FAQ
What is LongCat-2.0?
Meituan's LongCat team's frontier model: a 1.6-trillion-parameter mixture-of-experts with roughly 48 billion active parameters, a 1-million-token context window, and an architecture built for agentic coding. It is the full model behind the Owl Alpha stealth release on OpenRouter.
Is it really free?
In Nous Portal, yes — for one week. After that, the OpenRouter route it ran on as Owl Alpha remains the practical way in.
Does it work with the Hermes agent?
Yes — Hermes is provider-agnostic, so LongCat-2.0 is a model profile like any other. Your persona, skills, memory and schedules stay exactly as they are while the brain changes underneath.
Is it better than GPT-5.5?
On the lab's own SWE-bench Pro figures it edges ahead — 59.5 against 58.6. That is one benchmark, self-reported. The free week exists so you can answer the question on your own tasks instead of trusting anyone's chart, including theirs.
What should I test first?
A real coding task, one genuine long-context job, and an agent loop with tool calls — judged side by side against your current daily brain.
The Verdict
Free weeks on frontier models do not come around often, and this one lands in the exact ecosystem your agent already lives in. The model is credible, the shape is right for agent work, and the price for seven days is zero — which means the only real cost is the hour it takes to run your own tests. Spend the hour. Inside my Agent OS the swap took one profile change, and that is the whole point of building on an agent instead of a chat window: when a better brain shows up, you upgrade the thinking without touching the system.
If you want a setup where every new model is a one-click swap instead of a rebuild, check out the AI Profit Boardroom — the Agent OS with model profiles ready to go, daily tutorials that cover drops like LongCat-2.0 as they land, four live coaching calls a week and 3,700+ business owners building alongside you. → Run the new brains the day they drop
