Tencent open-sources Hy4, a 770B model that helped tune itself

Tencent has released and open-sourced Hy4 preview, a large mixture-of-experts model with 770B total parameters and 49B active per token, plus a context window above 1M tokens. Tencent aims it at coding, office work and scientific research. In an internal blind evaluation with 163 experts scoring 203 engineering tasks, Hy4 preview reached 2.99 out of 4.00, edging out GLM-5.3 at 2.92 and Kimi K3 at 2.94. Weights are open, and the model is available through Tencent Cloud TokenHub and OpenRouter, with free access for the first two weeks. Listed pricing is $0.834 per million input tokens and $2.501 per million output tokens, with cache hits at $0.042.

The detail worth pausing on is how it was built. Tencent says Hy4 took part in its own optimization, forming an early recursive self-improvement loop in which the model helped tune its inference system and pushed throughput up by 31.8%. That claim is narrow, it covers serving efficiency rather than the model retraining itself, but it is a concrete example of a lab handing an internal engineering task to the model it is shipping.

Why it matters

Another capable open-weight model at this scale gives teams a self-hostable option that scores near the top Chinese labs on blind engineering tasks, with pricing you can check before committing. If you evaluate open models, the two-week free window is a cheap chance to test Hy4 on your own workload rather than trusting the vendor's benchmark.