You've watched everyone rave about GLM as a cheap, capable assistant. HY3 is the next step: Tencent's brand-new Hunyuan 3 — released this month, Apache 2.0, even cheaper — and unlike a plain chat model, it doesn't just talk. It runs multi-step agentic tasks, writes production frontend code, builds Excel financial models with live formulas, and drafts full slide decks. This is the native desktop app that puts all of it one click away on your Mac or PC. Free for a full month, no credit card.
┌─────────────────────────────────────────────┐
│ HY3 Desktop │
│ ● Assistant ● Agent ● Code ● Docs │
│ │
│ "Build a 3D spinning Earth in one HTML file" │
│ → done. 1 file. Opens in your browser. │
│ │
│ "Model 3-region cash flow, 5000+ cells" │
│ → live Excel, every number a formula. │
└─────────────────────────────────────────────┘
No terminal, no config, no API key to start.
- Windows → download
HY3-x64.7zfrom RELEASES, double-click. Signed, passes SmartScreen. - Mac → download
HY3.dmgfrom RELEASES , drag to Applications. Notarized, universal (Apple Silicon + Intel). Open it, and your month of free access starts immediately. That's the whole setup.
GLM made the case that a Chinese open model can be a genuinely good everyday assistant at a fraction of the price. HY3 takes that further on three fronts:
It's newer and cheaper. Hunyuan 3 shipped this month under Apache 2.0 — 295B mixture-of-experts, but only 21B parameters active per token, so it runs fast and prices low: roughly a fifth of what closed assistants cost, with a free month to start here.
It actually does things. Where a plain assistant answers, HY3 executes. Native tool-calling and agent workflows mean you can hand it a multi-step job — research this, pull that, build the file, check it — and it runs the whole chain, not just the first reply. Tencent's own deployments report task-resolution jumping to 90% on real office workflows.
It's a real coder. In a blind evaluation by 270 professional developers on their own work tasks, HY3 scored 2.67/4 against GLM-5.1's 2.51 — with its biggest edge in frontend, data, and CI/CD. That viral "spinning 3D Earth in one HTML file" demo? That's HY3's home turf.
As your everyday assistant — ask anything, switch between a fast direct mode and a deep-reasoning mode depending on how hard the question is. 256K context means you can paste an entire report, contract, or long thread and it holds all of it.
As an agent — "find the top 10 competitors, pull their pricing, put it in a table." HY3 chains the steps and hands you the finished artifact, using tools cleanly instead of hallucinating.
As a coder — full frontend components, whole-file HTML/JS builds, refactors, debugging. Strong across Cline, KiloCode, and Claude-Code-style loops (Tencent reports under 4% accuracy variance across scaffoldings — meaning it behaves consistently whatever tool you drive it from).
As an office powerhouse — this is where HY3 surprises people. It built a consolidated cash-flow model across three regions and 5,220 linked cells with zero hardcoded values — every figure a live formula. Given 101 SKUs of sales data, it produced 12 analysis tables and a 30-slide deck with 20 charts. Financial modeling, spreadsheets, decks, research reports — real deliverables, not just chat.
A single toggle changes how hard HY3 works:
- Direct — instant answers for simple things, lowest cost
- Low — light chain-of-thought for everyday reasoning
- High — deep step-by-step for hard math, complex code, tricky agent plans You control the speed/quality trade-off per task instead of paying for max reasoning on every "what's the weather."
| HY3 (this app) | GLM-5.2 | ChatGPT (GPT-5.5) | Claude (web) | |
|---|---|---|---|---|
| Native desktop app | Yes | Browser/app | Partial | Browser |
| Agentic task execution | Built-in | Partial | Partial | Partial |
| Context window | 256K | 200K | 256K | 200K+ |
| Coding (dev blind-eval) | 2.67/4 | 2.51/4 | — | — |
| Office/Excel/decks | Strong | Moderate | Moderate | Moderate |
| License | Apache 2.0 | Open-ish | Closed | Closed |
| Cost to run | ~⅕ of closed | Low | High | High |
| Free to start | 1 month | Trial varies | Subscription | Subscription |
Is it really free? Yes — a full month, no credit card, through our launch access. After the month, plug in your own Tencent Cloud or OpenRouter key; HY3's pricing is among the lowest anywhere (roughly 1 yuan per million input tokens), so most people spend pocket change. The app is free and open source permanently.
Is it safe to install? The client is MIT-licensed with source on GitHub, code-signed on Windows and notarized on Mac, SHA-256 checksums on every release. The model itself is Tencent's official Apache-2.0 Hunyuan 3. Grab it only from the Releases page here, not random mirrors.
How is HY3 different from just using GLM? GLM is a strong assistant; HY3 is a strong assistant that also executes agent workflows, codes at a level that beat GLM-5.1 in a developer blind-eval, and produces real office deliverables (spreadsheets, decks, financial models). Newer, Apache 2.0, cheaper to run, and native on your desktop rather than a browser tab.
Do I need a powerful computer? No. HY3 runs in Tencent's cloud; your machine is just the client. Any modern laptop is fine — the heavy lifting isn't local.
What's the catch with the agent mode? None, but supervise it early. Agentic execution means HY3 can take multi-step actions with tools — powerful, but you'll want to watch the first few runs on any new task type until you trust its plan, same as any agent.
Built on Tencent Hunyuan 3 (HY3), released July 2026 under Apache 2.0. Independent third-party desktop client — not affiliated with, endorsed by, or operated by Tencent. "HY3," "Hunyuan," "GLM," "ChatGPT," and "Claude" identify the technologies this app uses or compares to (nominative fair use). Free-month access is provided through launch arrangements; model usage after that follows Tencent/OpenRouter pricing. Benchmark and deployment figures reflect Tencent's published July 2026 results — verify against your own workload. MIT-licensed client — see LICENSE.
Made something great with the agent or shipped a build with it? A ⭐ helps others find it.