In August 2026 Tencent put its Hy3 large language model into WorkBuddy, the company's agentic AI workspace, and made it free to use worldwide through August 31. That window has now closed, and the model it advertised is no longer the headline — Tencent launched Hy4 Preview into the same product on August 28. Here's what WorkBuddy and Hy3 actually are, what the free period covered, and where you can still use the model.
Quick Answer: Tencent's free Hy3 access on WorkBuddy ran from July 6 to August 31, 2026, and has ended. Hy4 Preview took its place in WorkBuddy in late August. Hy3 itself is still available through the OpenRouter API and as open weights under Apache 2.0.
What changed since the August announcement
| What was true in early August | What's true now (October 2026) |
|---|---|
| Hy3 free access worldwide through August 31 | Window closed. WorkBuddy's Hy3 free quota no longer applies; paid or quota-limited access depends on the platform. |
| Hy3 was the featured model | Hunyuan Hy4 Preview launched August 28 in both the Chinese and international WorkBuddy, plus CodeBuddy. |
| Hy3 free via OpenRouter | OpenRouter still serves Hy3 as a paid API (per-token pricing). |
| Open weights under Apache 2.0 | Unchanged — you can still self-host it free. |
What WorkBuddy is
WorkBuddy is Tencent's agent workspace: you give it a task in plain language — research, data analysis, a document or presentation — and it plans and executes the steps, using built-in "skills" and domain experts rather than leaving you to assemble the workflow yourself. Think of it as Tencent's answer to the agent-style products from OpenAI, Anthropic and Google, with WeChat-ecosystem roots in China and an international edition that went global in 2026.
One honest note on sourcing: the performance claims in Tencent's announcement — a reported 90%+ task success rate internally and a 68× jump in invocations over Hy2 — came from Tencent's own evaluations, not independent benchmarks. Treat them as the vendor's numbers.
The model picker (shown above) also makes the multi-model story plain: alongside Hy3 you'll find third-party options like Kimi-K3, with "Fast" through "Deep" reasoning-effort modes priced as rate multipliers.
Hy3 in numbers
The claims that matter for developers are the technical ones, which check out against the model cards:
- 295 billion total parameters, 21 billion activated — a Mixture-of-Experts design that keeps inference cost near a 21B model while carrying a 295B-weight memory.
- 256K-token context window — long documents and multi-step agent sessions in one context.
- Open weights, Apache 2.0 license — published on GitHub, Hugging Face, ModelScope and GitCode since July 2026, so you can download, fine-tune and self-host it with no per-token cost.
- OpenRouter API pricing — around $0.13 per million input tokens and $0.53 per million output tokens, which undercuts most frontier APIs by a wide margin for agentic workloads.
For agent tasks specifically — tool use, multi-step planning, code execution — that combination is the point: near-frontier reasoning at commodity prices.
How you can still use Hy3
1. Self-host the open weights. The Apache 2.0 release means the free option didn't actually end on August 31 — it just moved to your own hardware or cloud account. That's the genuinely free route.
2. Call it through OpenRouter. Pay per token as above, swap the model name in your existing OpenAI-compatible client, no sign-up with Tencent required.
3. Use WorkBuddy directly. The workspace continues; Hy3 remains selectable in the international edition, and Hy4 Preview is now the featured model. Availability and quotas vary by region — check the app itself rather than the August announcement. (The workbuddy.ai domain referenced in August was unreachable when this guide was last checked; start from Tencent's Hunyuan pages.)
Is "free agentic AI access" still a thing?
Yes, in two forms. First, the open-weights release: self-hosting Hy3 costs compute, not licence fees, and smaller deployments fit comfortably on a single high-memory GPU. Second, free tiers and preview windows — Tencent ran one for Hy3, and Hy4's preview launch came with its own promotional access. These windows are marketing; the durable free option is always the open weights.
FAQ
What is Tencent WorkBuddy?
It's Tencent's agentic AI workspace, available in Chinese and international editions. You assign it tasks in natural language — research, documents, data analysis — and it executes with built-in skills and a choice of models, including Tencent's own Hunyuan models and third-party options.
Is Hy3 still free on WorkBuddy?
No. The worldwide free-access window ran from July 6 to August 31, 2026, and has closed. Hy3 remains available through paid API access and, free of licence fees, as open weights you can self-host.
What replaced Hy3 as Tencent's flagship?
Hunyuan Hy4 Preview, launched August 28, 2026, in WorkBuddy and CodeBuddy. Demand at launch reportedly exhausted capacity in the early hours. Hy4 is a preview, so production users mostly stayed on Hy3 initially.
Can I run Hy3 myself?
Yes. Tencent released the weights under Apache 2.0 on GitHub, Hugging Face, ModelScope and GitCode. You'll need GPU hardware (or a cloud GPU) sized for a 295B MoE model — about 21 billion parameters activate per token, which keeps inference more affordable than a dense model of the same size.
