WorkBuddy users can now pick Tencent’s new flagship model from the conversation menu. Hy4 preview, the 770-billion-parameter open-weight Mixture-of-Experts model that Tencent published on 28 August 2026, went live inside the Chinese and international versions of WorkBuddy on the same day, and it is free to use there for two weeks, until 10 September.
That matters because WorkBuddy is not a chat window. It is Tencent’s desktop AI agent for office and coding work: you describe a task in a sentence, it plans the steps, works through your files, and hands back a deliverable such as a report, a spreadsheet, a slide deck or a working front-end. Putting a model with a one-million-token context window behind that loop changes what a single prompt can cover.
We covered the model itself, from its 49B active parameters to its Apache 2.0 licence and price list, in our Hy4 preview breakdown. This article is about the product side: what the integration changes for daily work, how the model switch works and what it is recommended for, the blind test Tencent ran in the WorkBuddy environment, the quota and reasoning-time caveats, and the routes for teams that would rather call the model directly or run it locally. Every figure traces to a link in the References section.
Table of contents
- What Tencent Announced
- How to Use Hy4 preview Inside WorkBuddy
- WorkBuddy Work Tasks vs Coding Tasks
- What WorkBuddy Actually Is
- The Blind Test Behind the Claim
- Hy4 preview Benchmarks That Matter for Agent Users
- Running Hy4 preview Outside WorkBuddy: API and Local GGUF
- WorkBuddy Against Claude Cowork and Other Agent Desktops
- Caveats Before You Move Work Into the Agent
- What the WorkBuddy Free Window Means for UK Teams
- WorkBuddy and Hy4 preview: Frequently Asked Questions
- References and Further Reading
What Tencent Announced
The corporate announcement is short on the product detail, so it is worth separating what was stated from what has since been documented by Tencent Cloud’s developer team.
The official line
Tencent’s newsroom post says Hy4 preview “will be available for free on WorkBuddy and CodeBuddy for two weeks” from launch, and that free access to the previous generation, Hy3, has been extended on both platforms until 30 September. The model is also live in Yuanbao and ima, Tencent’s consumer assistant and knowledge tool, and reachable through the Tencent Cloud TokenHub API and OpenRouter.
The product detail
A Tencent Cloud developer article published the next day fills in the mechanics. Hy4 preview appears in the app’s per-conversation model dropdown, tagged as limited-time free. Hy3 remains free and stays the default for quick tasks. The model choice applies only to the conversation where you make it, so switching in one thread does not change the others.
Why the co-design claim matters
Tencent says the model was trained on data co-created with its internal experts in software engineering, gaming, finance and security, and that it went through “deep co-design with products such as WorkBuddy”. In plain terms, the training and evaluation loops ran against the kinds of multi-step office and engineering tasks the agent actually executes, rather than against chat transcripts alone. That is the strongest argument for trying the model inside the agent first rather than through a raw API.
| Where Hy4 preview is available | Type | Cost at launch |
|---|---|---|
| WorkBuddy (China and international) | Desktop AI agent for office and coding work | Free for two weeks, then credits |
| CodeBuddy | IDE and CLI coding assistant | Free for two weeks, then credits |
| Yuanbao and ima | Consumer assistant and knowledge tool | Included |
| Tencent Cloud TokenHub | API | Metered per token |
| OpenRouter | API | $0.834 in / $2.501 out per million tokens |
| Hugging Face, GitHub, ModelScope | Open weights, Apache 2.0 | Free to download and self-host |
How to Use Hy4 preview Inside WorkBuddy
The switch takes a few seconds, but the guidance around it is worth reading before you commit a long task to the new model.
Switching the model
Open a conversation, find the model selector in the upper-right corner or beside the input box, and choose Hy4 preview from the list. The limited-time free label sits next to it. Because the setting is per conversation, a sensible pattern is to keep Hy3 on quick, routine threads and start a fresh conversation on Hy4 preview for anything heavy.
What Tencent recommends it for
The developer guidance names four task shapes where the new model is worth the extra reasoning time: lengthy document summaries, complete site code generation, formatted presentations, and cross-file analysis. These are exactly the tasks where a million-token window pays off, because the agent can load a whole folder of source material into one context rather than chunking it.
Quotas and reasoning modes
Two practical constraints apply. First, daily quota limits are in place during the free window, and going beyond them consumes credits from your plan. Tencent has not published the daily figure. Second, the model defaults to a high reasoning effort; if you want faster answers on simpler prompts, the API accepts a reasoning_effort setting of no_think, and Tencent’s own guidance warns that Hy4 preview can queue and over-verify on busy days.
What the model cannot do yet
Hy4 preview has no image or video input. Any task that depends on reading screenshots, scanned PDFs or design mock-ups still needs a different model in the switcher.
WorkBuddy Work Tasks vs Coding Tasks
Tencent markets the integration as covering both work and coding tasks, and the distinction is real: the two sides of the product use the model differently.
Work tasks
On the office side, WorkBuddy reads and writes native Word, Excel and PowerPoint files, runs data analysis, and can operate in a deep-research mode that gathers sources before drafting. Hy4 preview’s long context lets one conversation hold, say, a year of board papers or an entire supplier folder, and its recommended use cases in the product are precisely the document-heavy jobs.
Coding tasks
On the engineering side, the same agent can generate a complete site from a brief, work across a multi-file codebase, and hand deep repository work to its sibling, CodeBuddy, which shares the same architecture, billing and security audit. Tencent says the model brings “stronger understanding, planning, debugging, and validation capabilities for long-context development tasks”, and its Terminal-Bench 2.1 score of 85.4, up from Hy3’s 70.8, is the figure that best reflects that kind of work.
Where each one belongs
| Task | Best fit in the Tencent stack | Why |
|---|---|---|
| Summarise a 300-page tender pack | WorkBuddy on Hy4 preview | Whole pack fits one context |
| Reformat a monthly report into slides | WorkBuddy on Hy3 | Quick, formatted output; free default |
| Build a marketing site from a brief | WorkBuddy on Hy4 preview | Named use case: complete site code generation |
| Refactor a large repository | CodeBuddy on Hy4 preview | IDE and CLI tooling, same model |
| Analyse spreadsheets across folders | WorkBuddy on Hy4 preview | Named use case: cross-file analysis |
| Read screenshots or scanned documents | Another model in the switcher | Hy4 preview has no vision input |
What WorkBuddy Actually Is
For readers outside China the product may be new, so a short profile helps.
Origins and launch history
WorkBuddy was built by the Tencent Cloud CodeBuddy team and launched in China on 9 March 2026 after a test with 2,000 Tencent employees. It arrived with compatibility for OpenClaw skills, Model Context Protocol support, more than 20 skill packs, and one-click switching among Hunyuan, DeepSeek, GLM, Kimi and MiniMax models. The international version followed on 29 May 2026 at Tencent Cloud Day in Hong Kong.
Platforms and integrations
WorkBuddy runs as a desktop app on Windows 10 and 11 and on macOS 10.15 or later, with an iOS companion app. It can be driven remotely from Slack, Telegram, Discord, WeChat, WeCom, QQ, Feishu and DingTalk, and it ships MCP connectors for GitHub, Jira, Google Drive, Gmail, Notion and Slack. Custom OpenAI-compatible endpoints, including a local Ollama server, can be added as models.
Plans and pricing
| WorkBuddy plan (international) | Price | Credits |
|---|---|---|
| Free | $0 | 500 per month |
| Pro | $19.90 per month, or about $9.95 on annual billing | 1,000 per month |
| Team | $40 per seat per month | 1,000 per seat per month |
| TokenPlan Lite / Standard / Pro / Max | $7 / $17 / $51 / $103 | 1,000 / 2,600 / 7,900 / 15,900 |
Tencent has not published user numbers for WorkBuddy. The only proxies it offers are internal: daily token consumption in the app grew twenty-fold between the Hy3 preview and Hy3 general availability, and the share of users choosing Hy3 rose six-fold.
The Blind Test Behind the Claim
The most product-specific number in the release is a blind evaluation Tencent ran in the agent’s own environment, and it deserves a careful reading.
How it was run
Tencent had 163 of its own experts rate model outputs on 203 engineering tasks, scoring each on a four-point scale without knowing which model produced it. Hy4 preview averaged 2.99 out of 4.00. GLM 5.3 scored 2.92 and Kimi K3 scored 2.94.
What the margins really say
Against GLM 5.3, Hy4 preview won 46.8% of comparisons, tied 12.8% and lost 40.4%. Against Kimi K3 it won 51.2% and lost 40.9%. Tencent’s own phrasing is “slightly ahead”, and that is the honest reading: a 0.05-point lead on a four-point scale, with losses on roughly four tasks in ten, is a narrow win, not a rout. It is also an internal test with no independent replication yet.
Why it matters for the product anyway
Both rivals are models the app already hosts in its switcher, so the comparison is directly usable: if Tencent’s numbers hold, the free house model is now at least level with the paid alternatives on the agent’s own task mix.
Hy4 preview Benchmarks That Matter for Agent Users
The full benchmark sheet is in our model breakdown; these are the rows that map onto office and coding work. All are Tencent-reported.
| Benchmark | Hy4 preview | Relevance to WorkBuddy |
|---|---|---|
| Terminal-Bench 2.1 | 85.4 (Hy3: 70.8) | Multi-step coding in a shell |
| SWE-bench Multilingual | 82.9 | Fixing real repository issues |
| SWE-Bench Pro (public) | 65.7 | Harder, longer engineering tickets |
| MCP-Atlas | 83.7 | Using MCP tools, which WorkBuddy relies on |
| Toolathlon-Verified | 74.1 | Long tool-use chains |
| APEX-Agents (pass@1) | 37.1 (Kimi K3: 37.2) | One-shot agent success |
| GPQA Diamond | 92.3 | Graduate-level reasoning |
The one that moved most
A 14.6-point rise on Terminal-Bench is the clearest signal that the coding side of the product gets the bigger upgrade. The office-side proxies are softer: Tencent cites an internal task-completion rate above 90% in office scenarios and a 34% reduction in completion time against the previous generation, neither of which has been verified outside the company.
The self-optimisation footnote
Tencent also says Hy4 preview took part in optimising its own training methods, data strategy and low-level operators, producing a 31.8% inference-throughput gain. Read that as a description of the build process, in which the model proposed experiments and engineers fed the results back, rather than as a claim of autonomy. The post-training itself leaned on reinforcement learning against expert-built task data, which is the more conventional source of the gains.
Running Hy4 preview Outside WorkBuddy: API and Local GGUF
Not every team wants a desktop agent from a Chinese cloud provider handling its files. The same model is available two other ways, and a third arrived two days after launch.
The API route
On OpenRouter the model lists a 1,048,576-token context, a 64,000-token completion cap, function calling and JSON-schema structured output, at $0.834 per million input tokens and $2.501 per million output tokens, with cached input at $0.042. That undercuts the two rivals in the blind test: GLM-5.3 lists at $1.40 in and $4.40 out, Kimi K3 at $3 and $15. DeepSeek V4 Pro is cheaper still at roughly $0.44 in and $0.87 out.
The local GGUF route
On 29 August Tencent’s Hy account announced that it had compressed Hy4 preview “from 1.5TB to about 200GiB GGUF and it still works well”. The AngelSlim team published two llama.cpp builds on Hugging Face: a Q4_K_M file at 435.20 GiB and a mixed-precision STQ1_0 file at 213.66 GiB, the latter averaging 1.3125 bits per weight with calibration data choosing each layer’s bit-width. On eight H20 GPUs with the weights fully resident, the STQ1_0 build decoded at 20.47 tokens per second and prefilled at 204.56 tokens per second.
What that means in practice
| Route | What you need | Best for |
|---|---|---|
| WorkBuddy app | Windows or macOS, free account | Office deliverables and light code with no setup |
| OpenRouter or TokenHub API | An API key and your own tooling | Integrating the model into existing workflows |
| GGUF Q4_K_M | About 435 GiB of GPU memory, patched llama.cpp | Highest local fidelity |
| GGUF STQ1_0 | About 214 GiB of GPU memory, local disk, CUDA 13 | Data-sovereign deployment on one node |
| Full FP8 weights with vLLM or SGLang | Eight-GPU server | Production self-hosting |
Even the smallest build is a server-class job; it is not a laptop model. The point of the compression is that a single node, rather than a rack, can now hold the full 770B model, which brings self-hosting within reach of a mid-sized firm with a compliance reason to keep documents on-premises. Keep the file on local disk: the AngelSlim notes warn that memory-mapping the weights over NFS runs at around 12 MB/s on random page faults and can turn a one-minute load into hours.
WorkBuddy Against Claude Cowork and Other Agent Desktops
WorkBuddy’s most direct Western comparison is Anthropic’s Claude Cowork, and the two products make different trade-offs.
| Feature | WorkBuddy | Claude Cowork |
|---|---|---|
| Free tier | Yes, 500 credits a month | No; about $20 Pro, about $25 per user Team |
| Model choice | Hy4 preview, Hy3, GLM, Kimi, DeepSeek, MiniMax, custom endpoints | Claude Opus and Sonnet only |
| Native Office file writing | Word, Excel, PowerPoint | Via file tools |
| Remote control | Slack, Telegram, Discord, WeChat, Feishu, DingTalk and more | Narrower |
| English writing and reasoning | Good, improving with Hy4 preview | Generally stronger |
| Data residency | Tencent Cloud | Anthropic |
The Chinese field
Inside China the comparison set is the agents from ByteDance, Alibaba and Zhipu, but no published test pits them against WorkBuddy on Hy4 preview yet. What Tencent’s switcher does show is confidence: it hosts GLM-5.3, Kimi K3, DeepSeek V4 and MiniMax M3 alongside its own model, then benchmarks itself against them.
The coding-agent question
Against Claude Code and Codex, the sensible framing is complementary rather than head-to-head. WorkBuddy is built for office deliverables and light code; repository-native refactors belong with a terminal-first agent, or with CodeBuddy inside the Tencent stack. For more on how that category is evolving, see our coverage of AI employees and autonomous AI agents.
Caveats Before You Move Work Into the Agent
A free two-week window is an invitation to test, not a reason to migrate. Five things should temper the enthusiasm.
Tencent’s own disclosures
The model card describes Hy4 preview as an early version with known issues: it “tends to spend longer than necessary reasoning through complex tasks” and “over-verifies its own work”. Expect slower turnarounds than Hy3 on simple prompts, and budget for it.
The hands-on record is mixed
An InfoQ China reporter ran the model through WorkBuddy on launch day. A procurement report went wrong on price arithmetic; a four-platform dashboard build passed its own 147 tests but shipped with path errors; a quarterly business review needed two rounds of Word-formatting fixes. The verdict was that the agent shows the delivery ability of a small team, but someone still has to watch it.
Every benchmark is vendor-reported
Nothing above has an independent leaderboard result yet. Early third-party monitoring of the OpenRouter endpoint recorded roughly 86% uptime, a median of 36 tokens per second and 3.19 seconds of latency across the first three days, which is a sign of demand more than of stability.
Data and jurisdiction
WorkBuddy processes files through Tencent Cloud. For a UK firm handling personal data or regulated material, that is a data-transfer question to settle before the trial, not after. Teams with a hard residency requirement have the GGUF and vLLM routes instead.
The window closes
Free Hy4 preview use ends on 10 September 2026. After that the model bills against your credits, and the daily quota already applies during the free period.
What the WorkBuddy Free Window Means for UK Teams
For a small or mid-sized business the practical question is not whether the model wins a benchmark but whether two free weeks can answer a real question about your own work.
A test worth running
Pick three recurring jobs: one document-heavy, one spreadsheet-heavy, one small coding task. Run each through the agent on Hy3 and then on Hy4 preview, and score the output the way the InfoQ reporter did, on whether a colleague could ship it without a rewrite. That is a better signal than any blind-test average.
Cost implications
If the test lands well, the numbers are attractive: the free tier’s 500 credits, a $19.90 Pro plan, and API pricing under a dollar per million input tokens. If it does not, you have lost a fortnight and learned where the current generation of agents still needs a supervisor.
Where to go next
The full model breakdown, including architecture, serving recipes and the complete benchmark table, is in our Hy4 preview article, and this month’s other open-weight releases, including Alibaba’s Qwen3.8-Flash preview and Z.ai’s GLM-5.3-Flash, sit in our AI models and tools hub. If you want help deciding whether an agent desktop belongs in your workflow at all, our AI consulting team runs exactly this kind of structured trial.
WorkBuddy and Hy4 preview: Frequently Asked Questions
Is Hy4 preview free on WorkBuddy?
Yes, for two weeks from 28 August 2026, so until 10 September. Daily quota limits apply, and use beyond them draws on your credits. Hy3 stays free on both WorkBuddy and CodeBuddy until 30 September.
Do I have to switch to Hy4 preview?
No. Hy3 remains the default for quick tasks. Hy4 preview is a per-conversation choice in the model selector.
Is WorkBuddy available in the UK?
Yes. The international version launched on 29 May 2026 and runs on Windows and macOS, with an iOS app and a free tier of 500 credits a month.
Can WorkBuddy handle both work and coding tasks?
Yes. It writes native Word, Excel and PowerPoint files, runs analysis and research, and generates code, including complete sites. Deep repository work is better handled by CodeBuddy, which shares the same model and billing.
What is the difference between WorkBuddy and CodeBuddy?
Both come from the Tencent Cloud CodeBuddy team and share an architecture. CodeBuddy is the IDE and command-line coding assistant; WorkBuddy is the desktop agent for general office work that also covers lighter coding.
Can I run Hy4 preview locally instead?
Yes, if you have server-class hardware. The GGUF builds need roughly 214 GiB or 435 GiB of GPU memory with a patched llama.cpp, and the full FP8 weights run on an eight-GPU node with vLLM or SGLang.
References and Further Reading
Tencent: Tencent Releases and Open-Sources Tencent Hy4 preview
Tencent Cloud Developer: Using Hy4 preview in WorkBuddy
Tencent Cloud Developer: Hy4 preview release and two-week free window in WorkBuddy
WorkBuddy: Product overview and documentation
Hugging Face: tencent/Hy4-preview model card
GitHub: Tencent-Hunyuan/Hy4-preview
Hugging Face: AngelSlim/Hy4-preview-GGUF quantised builds
Tencent Hy on X: Hy4-preview compressed from 1.5TB to about 200GiB GGUF
OpenRouter: Hy4 preview pricing and providers
AIbase: Tencent WorkBuddy launches integration with Hy4 preview
TechNode: Tencent open-sources Hy4 preview with 770B parameters and a 1M-token context
TechNode: Tencent launches WorkBuddy productivity AI agent for global users
TechNode: Tencent launches OpenClaw-like workplace AI agent WorkBuddy
InfoQ China: Hands-on with Hy4 preview inside WorkBuddy
BigGo Finance: Tencent Hunyuan releases Hy4 preview across multiple products
MindStudio: Inside the Tencent Hy4 preview open-weight model
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.