Tencent released Hy4 Preview on August 28, 2026 — a 770-billion-parameter Mixture-of-Experts AI model licensed under Apache 2.0 and available via API at $0.834 per million input tokens through OpenRouter. With 49 billion active parameters per inference pass and a context window exceeding 1 million tokens, Hy4 Preview is now one of the most cost-efficient frontier-class models accessible globally.
Tencent Hy4 Preview is a 770B open-weight MoE model at $0.834/M input tokens — below GPT-5.6 Sol ($4/M) and Gemini 3.7 Flash ($1.50/M). Licensed Apache 2.0 with no royalty restrictions, it supports 1M+ token context and is accessible globally via OpenRouter without a Tencent Cloud account.
What Is Tencent Hy4 Preview?
Tencent Hy4 Preview is a 770-billion-parameter Mixture-of-Experts language model released by Tencent’s HunYuan division on August 28, 2026, under an Apache 2.0 open-source licence. The model activates 49 billion of its 770 billion parameters per inference pass, which keeps per-token compute cost at the level of a 49B dense model while drawing on the full 770B parameter space for output quality.
Hy4 Preview supports a context window exceeding 1 million tokens, enabling tasks that require processing large volumes of text in a single pass — such as full-contract legal review, large-codebase comprehension, and multi-document synthesis. It follows Tencent Hy3, its 295B predecessor, which also carried an Apache 2.0 licence and used the same MoE architecture.
How Hy4 Preview Compares to Kimi K3 and GLM 5.3
In Tencent’s internal blind evaluation — conducted by 163 domain experts across 203 engineering tasks — Hy4 Preview scored 2.99 out of 4, placing ahead of Alibaba’s Kimi K3 at 2.94 and Zhipu’s GLM 5.3 at 2.92. The evaluation covered coding, data analysis, and office automation tasks judged without model attribution disclosed to the assessors.
Separately, infrastructure improvements built during Hy4’s development increased Tencent’s own end-to-end training and inference throughput by 31.8% compared to the prior baseline — an efficiency gain independent of the cross-model benchmark results.
Tencent’s benchmark did not include GPT-5.6 Sol or Gemini on any Western standardised evaluations. All cross-model scores above are from Tencent’s internal engineering-task assessment only.
Pricing and Access: $0.834/M Input Tokens via OpenRouter
Hy4 Preview costs $0.834 per million input tokens and $2.501 per million output tokens, available via Tencent Cloud TokenHub and OpenRouter. For pricing context: GPT-5.6 Sol’s input cost is $4 per million tokens as of August 2026, and Gemini 3.7 Flash is approximately $1.50 per million tokens — Hy4 Preview’s input price is below both.
Free access is available for a limited two-week period through Tencent’s WorkBuddy enterprise productivity suite and CodeBuddy developer environment. Additional access points include the Yuanbao and ima applications, plus Tencent Cloud TokenHub for API integration. OpenRouter provides the globally accessible API endpoint for teams based outside China.
The “preview” designation means these pricing and performance specifications may change at general availability. Tencent has not announced a GA date for Hy4.
What Hy4 Preview Means for Business Teams
Business teams evaluating AI tools for coding assistance, high-volume document processing, and office automation now have a frontier-class option at sub-$1/M input pricing under Apache 2.0 terms — meaning no royalty restrictions on commercial deployment. The OpenRouter endpoint makes Hy4 accessible for international organisations without requiring a Tencent Cloud account.
Chinese open-source labs — Tencent, Alibaba, and Zhipu — are now releasing Apache-licensed large models with pricing well below $1/M input, which applies direct competitive pressure on OpenAI and Anthropic’s enterprise pricing tiers. For a broader view of which AI tools are currently leading for business use cases, see our guide to the best AI tools for business in 2026.
For Context: Tencent’s HunYuan division previously released Hy3 (295B parameters, Apache 2.0) earlier in 2026. Hy4 Preview represents a 2.6× scale increase in total parameter count while maintaining the same open-licence model.
Related: Best AI agents for business tasks in 2026 — see how Hy4-powered tools like WorkBuddy and CodeBuddy sit within the broader enterprise agent landscape.
