Tencent: Hy4

Tencent released Hy4 preview on 28 August 2026 with 770 billion backbone parameters, 49 billion active per token, and a serving context window above one million tokens, with weights and code under Apache 2.0. Tencent explicitly labels this an early version that can reason and verify longer than necessary, which is why the reviewer recommends it only for long-context, multi-step tool-use coding work rather than as a universal production default. A hands-on WorkBuddy coding-agent test led the author to propose a two-week pilot for teams constrained by long context, repeated repository retrieval, and multi-step tool use. Because the checkpoint is enormous, self-hosting is framed as a datacenter-scale project, and the piece cautions that a large context window does not by itself prove reliable recall.

Tencent's Hy Team released Hy4 preview on 28 August 2026 as an open-weight Mixture-of-Experts model with 770 billion total parameters and 49 billion activated per token, distributed under Apache License 2.0. Weights ship in two forms — a full-precision Hy4-preview checkpoint and an FP8-quantized Hy4-preview-FP8 checkpoint — published on Hugging Face, ModelScope, GitCode, and CNB. The model supports roughly a one-million-token context window and includes a built-in multi-token prediction layer for speculative decoding. Tencent's blind comparative evaluation was run internally, so the published benchmark standing should be read as a snapshot of a self-selected internal test suite rather than a third-party audited ranking, and the model's position relative to rivals remains unsettled.