What is Tencent’s Hy4 Preview open-source AI model, and how does it compare with its Hy3 predecessor and Chinese rivals in terms of parameter count, activated parameters, context window, benchmark performance, availability, integrations, pricing, limitations, expected improvement

Hy4 Preview is Tencent’s newly open-sourced, mixture-of-experts (MoE) flagship aimed at agentic productivity—particularly software engineering, research, office automation, and tool use. It is a major scale-up from Hy3, but it remains a preview release: Tencent’s strongest comparative results are chiefly vendor-reported, so independent validation is still limited. 414

Hy4 Preview versus Hy3

Metric	Hy3 Preview	Hy4 Preview	Meaning
Total parameters	295B	770B	2.6× larger overall MoE capacity. 24
Activated parameters	~21B/token	~49B/request	Hy4 uses roughly 2.3× more active compute, so it is likely more capable but materially more expensive to serve. 24
Context window	256K tokens	>1M tokens	Hy4 offers about four times the nominal context, useful for large codebases, long documents, and multi-step agents. 214
Intended use	Reasoning, coding, RAG/IDE agents	Agents, coding, research, office workflows, complex tool use	Hy4 is positioned more explicitly as a production productivity-agent model. 24
Release status	Open-weight model; reported Apache 2.0 terms	Open-sourced preview	Hy4’s exact licence and self-hosting terms should be checked in its official repository before commercial deployment. 214

Hy3 was already unusually compute-efficient for its size: its MoE routes roughly 21B of 295B parameters per token, using 192 experts with top-8 routing; it also had configurable reasoning depth and reported 74.4% on SWE-bench Verified. 23 Hy4 preserves the sparse-MoE economic logic but increases both total capacity and inference compute substantially. 414

Performance and Chinese-rival positioning

• Tencent reports 85.4 on Terminal-Bench 2.1, a 14.6-point gain over Hy3, and says Hy4 exceeds DeepSeek V4 Pro on that terminal-programming evaluation while matching Claude Opus 5. This is a Tencent-associated claim rather than an independently reproduced ranking. 13
• In Tencent’s blinded expert evaluation of 203 engineering tasks, Hy4 averaged 2.99/4, versus 2.92 for Z.ai’s GLM-5.3 and 2.91 for Moonshot’s Kimi K3. This is a narrow result and should not be treated as a universal leaderboard. 5
• By raw total parameters, Hy4’s 770B is smaller than reported DeepSeek V4 Pro at 1.6T and Kimi K3 at 2.8T, while close to GLM-5.3 at 744B. Its 49B activated count is reported to be broadly comparable with DeepSeek V4 Pro. 4
• For context, an earlier GLM-5 result was reported at 77.8% SWE-bench Verified. Cross-model claims need care: benchmark versions, tool settings, test contamination controls, and agent scaffolding can differ. 1

The practical conclusion is that Hy4 looks competitive among Chinese frontier open-weight models in coding and agent workflows, but there is insufficient independent evidence yet to declare it the overall leader over DeepSeek, Qwen, GLM, Kimi, or MiniMax.

Availability, integrations, and price

• Tencent has made Hy4 Preview available in the Chinese and international versions of WorkBuddy and CodeBuddy, providing a direct end-user route for workplace agents and coding agents. 14
• It is also open-sourced, which gives organizations a self-hosting option, although a 770B MoE model remains demanding to store and deploy even if only 49B parameters are activated per inference. 414
• Reported Chinese API pricing is RMB 6 per million input tokens, RMB 18 per million output tokens, and RMB 0.3 per million cached-input tokens. Pricing, quotas, regions, and API availability may change during preview. 4
• One third-party comparison lists Hy4 pricing at $0.85/M input and $0.62/M output, which conflicts with the reported RMB tariff and is internally counterintuitive for output pricing. I would rely on Tencent’s current console price rather than that aggregator. 84

Limitations and likely next steps

• Preview maturity: model quality, API behavior, rate limits, pricing, safety controls, and deployment documentation can change quickly. 14
• No apples-to-apples external proof yet: the most striking Hy4 comparisons come from Tencent’s own testing or launch reporting; independent, reproducible evaluations are needed. 513
• High deployment footprint: sparse activation reduces per-token compute, but serving a 770B model still requires substantial memory, distributed infrastructure, and operational expertise. Its active set is much larger than Hy3’s. 24
• Long context is not automatically reliable: a 1M-token window is valuable, but users should test retrieval accuracy, latency, and cost at their actual context lengths.
• The expected next improvements are therefore likely to be a non-preview/stable release, stronger agent/tool reliability, better inference efficiency and quantizations, independent benchmark validation, and broader cloud/API integrations. Those are reasonable product expectations, not announced commitments.

Tencent’s infrastructure bet and WorkBuddy traction

Tencent is pairing the model release with a sizeable capacity investment. In Q2 2026 it reported operating capital expenditure of RMB 51.8B, up 190% year over year, which contributed to negative RMB 13.8B free cash flow; management attributed the spending surge to AI infrastructure. 15

• Tencent stated that WorkBuddy was China’s most widely used productivity AI-agent service, while Bloomberg described it as China’s most popular AI workplace assistant shortly after its March launch. These are market-position claims, not audited revenue metrics. 71
• Third-party desktop-web traffic estimates put WorkBuddy at 20.97M June 2026 visits, up from 8.85M at launch in March; the same source explicitly cautions that visits are not daily active users, paying companies, revenue, or mobile use. 12
• Tencent management says both WorkBuddy and CodeBuddy are gaining users rapidly. 6
• The strategic logic is clear: Tencent can amortize GPU and cloud investment across its own assistants, Tencent Cloud customers, and its existing business ecosystem, rather than treating Hy4 as a standalone model release. The near-term tradeoff is higher capital intensity and uncertain monetization while products are still scaling. 615