Tencent releases Hy4 preview: open weights, 770 billion parameters, Apache 2.0 licence

On 28 August 2026 Tencent published the weights of “Tencent Hy4 preview”, making them available on Hugging Face in the main repository and in an FP8 quantised variant. The model uses a Mixture-of-Experts architecture with 770 billion total parameters, arranged over 78 layers: the first with a dense FFN, the remaining 77 with MoE, each holding 256 routed experts plus 1 shared. For every token the eight highest-scoring routed experts fire along with the shared one, which accounts for the 49 billion active parameters. The declared context window is 1 million tokens. What sets the release apart is the Apache 2.0 licence, with no community clauses and no caps on user volumes, alongside integration into company products such as CodeBuddy, WorkBuddy, Yuanbao and ima, plus API access on Tencent Cloud TokenHub (priced from $0.834 per million input tokens and $2.501 per million output tokens) and OpenRouter.

The official card reports scores of 82.9% on SWE-Bench Multilingual, 65.7 on SWE-Bench Pro, 85.4 on Terminal-Bench 2.1, 64.3 on Deep SWE and 92.3 on GPQA Diamond. The company has also released the results of a blind evaluation run in-house by 163 experts across 203 engineering tasks, where the model averaged 2.99 out of 4.00 against 2.92 for GLM-5.3 and 2.94 for Kimi K3 — scores likewise assigned by Tencent, not by an outside evaluator. The documentation openly acknowledges some operational limits of the architecture: the model “can sometimes take longer than necessary to work through complex questions and may over-verify its own answers”, a passage picked up by Reuters as well. In the model card the company adds that there is “real headroom left in both pre-training and post-training”.

Every available metric traces back to Tencent's own evaluations and protocols: no independent audits or third-party reproductions have appeared. In the same way, the 31.8% gain in end-to-end throughput — credited to the model's autonomous optimisation of training and inference procedures — comes with no statement of the comparison baseline, and neither the training costs nor the exact make-up of the data co-built with in-house specialists are disclosed. The move places Tencent in the open-weights race alongside Alibaba, Z.ai and Moonshot; the open release sits beside monetisation through APIs and products, it does not replace it.

Releasing open weights with no restrictive clauses is a straightforward choice, but how solid the architecture really is can only be judged once the performance data leaves the circuit of in-house reports. — Olya