Tencent releases Hy4 preview, open-source AI model with self-improvement capabilities

Tencent has released the Hy4 preview, a 770-billion-parameter open-source Mixture-of-Experts AI model that marks a significant step in recursive self-improvement. Under human supervision, the model optimized its own training and inference systems, yielding a 31.8% throughput increase. While Goldman Sachs analysts praise Tencent's ecosystem-driven training strategy, security experts warn of pipeline vulnerabilities, and developers must navigate strict Chinese cybersecurity laws when using Tencent's API.

Hy4 preview release specifications

• ▪The Hy4 preview is released under the Apache 2.0 license, allowing developers to perform commercial use, modification, redistribution, and fine-tuning without royalties.
• ▪The Hy4 preview features a context window of 1,048,576 tokens, which is a four-fold increase over the 256,000-token context window of the previous Hy3 model.
• ▪Tencent released the Hy4 preview on August 28, 2026, as an open-source Mixture-of-Experts model featuring 770 billion total parameters and 49 billion active parameters.

Recursive self-improvement capabilities

• ▪The supervised recursive self-improvement loop of the Hy4 preview produced a 31.8% increase in end-to-end inference throughput by optimizing operator fusion and inter-GPU communication.
• ▪Tencent used the Hy4 preview to automate the optimization of its own training methods, data strategies, evaluation frameworks, and low-level system operators.

Ecosystem-driven training strategy

• ▪Tencent co-designed and validated the Hy4 preview using its internal WorkBuddy and CodeBuddy products, which are offering free access to the model through approximately September 11, 2026.
• ▪Goldman Sachs analysts stated that Tencent's product-plus-model strategy enabled it to collect user data from its product suite to feed back into subsequent training rounds.

Security implications for enterprises

• ▪The 2026 International AI Safety Report identifies uncontrolled recursive self-improvement as among the most consequential national-security-level risks associated with advanced AI.
• ▪A Cloud Security Alliance threat analysis warns that when AI participates in building its successor, the training pipeline becomes a high-value target where compromises could produce invisible behaviors.

Regulatory requirements under Chinese law

• ▪China's National Intelligence Law requires all organizations and citizens to cooperate with state intelligence on demand, which applies to developers routing API queries through Tencent Cloud.
• ▪Amendments to China's Cybersecurity Law effective January 1, 2026, explicitly extend the requirement to provide technical interfaces and decryption assistance to AI systems.

Performance benchmarks on Code Arena

• ▪The Hy4 preview ranked eighth globally on Code Arena's WebDev leaderboard on September 1, 2026, outperforming Alibaba's Qwen 3.8-Flash-Next in ninth place.
• ▪Tencent's previous generation model, Hy3, ranked 34th on the Code Arena WebDev benchmark as of September 1, 2026, highlighting the coding gains made by the Hy4 preview.