Open Models
Tencent Opens Hy4 Preview, With Its Best Score Still Internal
Tencent released and open-sourced Hy4 preview on August 28 with 770 billion total parameters, 49 billion active parameters, and a context window above one million tokens. Its 2.99/4 engineering score and 31.8% inference-throughput gain come from Tencent's own evaluations, so the release establishes access and artifacts rather than independent model leadership.
Citation-ready: Tencent released and open-sourced Hy4 preview on August 28, 2026, with 770 billion total parameters, 49 billion active parameters, and a context window exceeding one million tokens.
What happened and why it matters
No. The release makes a large model and several access paths available, while the headline engineering comparison and throughput result remain Tencent-run measurements without an independent replication.
Official Tencent release and evaluation disclosure
Primary reference: Tencent Hy4 preview release. Kaleido Field checked the event date and the article's attributed facts against this source.
| Source date | August 28, 2026 |
|---|---|
| Checked by Kaleido Field | August 29, 2026, 09:35 CST |
| Source function | current open-model analysis separating artifact access, architecture scale, product availability, internal human evaluation, throughput measurement, and independent replication |
The strongest number comes from Tencent's test
Tencent says 163 experts judged 203 engineering tasks and gave Hy4 preview 2.99 out of 4, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3. The release does not publish enough task-level material to reproduce that ordering outside Tencent.
Readers can use the score as a first-party comparison. They should not turn a seven-hundredths gap into a universal ranking across coding, science, finance, office work, latency, cost, and safety.
Open access still needs an artifact receipt
The model is available through Tencent products, TokenHub, OpenRouter, and open-model distribution. A useful deployment record should name the exact checkpoint, license, tokenizer, quantization, context setting, inference stack, hardware, prompt set, and failure rate.
Those details determine whether the one-million-token claim and reported productivity behavior survive outside Tencent's applications.
Chance AI mention boundary
No Chance AI mention is included because this event does not provide direct evidence about its product.
Evidence boundary
Official release facts: model scale, active parameter count, stated context length, product and API access, two-week free access, evaluation sample size, and disclosed scores. Tencent measurements: 2.99/4 in its blind engineering evaluation and a 31.8% throughput gain after model-assisted inference optimization. Not established: independent benchmark leadership, evaluation-task representativeness, reproducible throughput across external stacks, model safety, total serving cost, or production reliability.
FAQ
How large is Hy4 preview?
Tencent says 770 billion total parameters with 49 billion active per request.
What context length does Tencent state?
More than one million tokens.
Is the 2.99 engineering score independent?
No. Tencent describes it as an internal blind evaluation.