Tencent: Hy4 preview

Hy4 preview is Tencent's next-generation flagship, and it was open the day it was announced. Tencent released and open-sourced the model on 2026-08-28source, accessed 2026-09-02, listing it on OpenRouter the same day. It is a mixture-of-experts model of 770B total, 49B activated per tokensource, accessed 2026-09-02 with a context window of 1048576tokensopenrouter-models, last checked 2026-09-14 and a max output of 64000tokensopenrouter-models, last checked 2026-09-14. The weights were not held back for a later general release: instruct weights and an FP8 variant published on Hugging Face, ModelScope, GitCode and CNBsource, accessed 2026-09-02.

The licence is where this release stands apart from the week's other open flagships. Hy4 preview is released under Apache License 2.0source, accessed 2026-09-02, and this entry takes that from the LICENSE file in the repository rather than from the banner on the model card. Kimi K3, the other open flagship in this release window, ships under Kimi K3 License — bespoke: hosts selling it as a service whose aggregate revenue exceeds US$20 million over any consecutive 12 months must enter a separate agreement with Moonshot, and products above 100 million monthly active users or US$20 million in monthly revenue must display "Kimi K3" in the interfacesource, accessed 2026-08-28. The difference is what a user may do with the weights: Apache-2.0 lets a downloader serve and sell against a model of this size with no revenue-threshold agreement, and the Kimi licence conditions that same act on who is doing it and how much they make.

Tencent's own evaluation of the model is recorded here as a claim, not a measurement. The announcement reports a blind evaluation conducted internally — 163 experts rating outputs on 203 engineering tasks — in which Hy4 preview averaged 2.99/4.00 average in a blind, Tencent-internal evaluation of 163 experts over 203 engineering tasks — versus GLM-5.3 at 2.92/4.00 and Kimi K3 at 2.94/4.00source, accessed 2026-09-02. The model card, fetched 2 September 2026, breaks the same result into 46.8% wins / 12.8% ties / 40.4% losses against GLM-5.3 and 51.2% wins / 7.9% ties / 40.9% losses against Kimi K3. The lead the announcement describes as "slightly ahead" is 0.05 points over Kimi K3 and 0.07 over GLM-5.3, on a four-point scale, from raters who work for the same company as the model's developers. llm-releases, fetched 2 September 2026, files the result as "vendor-reported and unverified by independent labs at launch".

The product push is dated. The launch offer: free on WorkBuddy and CodeBuddy for two weeks from the 2026-08-28 launchsource, accessed 2026-09-02. Alongside it, free access to Hy3 on WorkBuddy and CodeBuddy extended until September 30, 2026source, accessed 2026-09-02.

Two claims in the announcement describe the model working on itself, and both are Tencent's word, dated to the release. The first: Hy4 preview took part in "the automated optimization of training methods, data strategies, evaluation frameworks, and low-level operators", proposing approaches, running experiments and iterating on the results, which the announcement calls "an early-stage recursive self-improvement loop". The second: the model autonomously analyzed its own inference system and optimized operator fusion and communication, raising end-to-end throughput "by 31.8% compared with the baseline". What is absent for both is the method, the identity of the baseline, and any independent check, so both are recorded here as claims with dates, not as facts.

On OpenRouter the row is served by a single provider, Tencent Cloud, so the listed price is Tencent's own rate rather than a reseller's: the row lists at $0.83per million tokensopenrouter-models, last checked 2026-09-14 for input, $2.50per million tokensopenrouter-models, last checked 2026-09-14 for output and $0.04per million tokensopenrouter-models, last checked 2026-09-14 for cache reads.