Tencent Open‑Sources 770‑B Hy4 LLM for Coding and Announces Self‑Improvement Loop

Tencent open‑sourced Hy4 preview, a 770‑billion‑parameter Mixture‑of‑Experts model with 49 billion active parameters per token and a context window beyond 1 million tokens, aimed at coding, office productivity, and scientific research. It is released under the Apache 2.0 license, a fully permissive open‑source license with no usage cap, no field‑of‑use restriction, and no commercial agreement required, unlike the restricted community licenses some rival open‑weight labs attach to their largest releases.

The release puts Tencent alongside other labs racing to ship massive, freely usable open‑weight models. Developers and enterprises can fine‑tune, redistribute, and deploy Hy4 commercially without negotiating terms with Tencent.

On benchmarks, Tencent's own internal blind evaluation, run across 163 engineers judging 203 engineering tasks, scored Hy4 preview at 2.99 out of 4.00, narrowly ahead of GLM 5.3 (2.92) and Kimi K3 (2.94). Tencent has not published independently audited benchmark results for Hy4.

Tencent also said Hy4 contributed to its own development: the model helped automate optimization of Tencent's training methods, data strategies, evaluation frameworks, and low‑level operators, and optimized its own inference stack, lifting reported throughput by 31.8%. Tencent describes this as an early‑stage recursive self‑improvement loop inside its training and research pipeline, not a live feature that revises outputs from end‑user feedback. The caution some coverage raised concerns governance of Tencent's model‑building process, not unpredictable runtime behavior in deployed code.

Analysts will watch developer adoption of the permissive license and whether the self‑improvement approach to training becomes a durable edge in future Hy4 releases.

Related Stocks

• TCEHY