Tencent

In nine days in August 2026 Tencent listed both the shortest context windows in the year's catalog and one of the longest. tencent/hy-mt2-7b, tencent/hy-mt2-30b-a3b and tencent/hy-mt2-1.8b, listed on 19 and 20 August, are translation models under Apache License 2.0, covering translation among 33 languagessource, accessed 2026-08-28, each advertising 8192tokensopenrouter-models, last checked 2026-09-24 — smaller than any other row listed in 2026, in a snapshot where the next smallest is exactly twice that. On 28 August the same vendor listed tencent/hy4-preview at 1048576tokensopenrouter-models, last checked 2026-09-24; it is 770B total, 49B active — mixture-of-expertssource, accessed 2026-08-28. That is a factor of 128 between two rows from one company inside a fortnight, and it is what a purpose-built model looks like next to a general one: a translator does not need to hold a codebase in mind.

Tencent's release pattern is preview first, then weights and a cheaper row. tencent/hy3-preview went up on 22 April 2026; tencent/hy3 followed on 6 July, heading at $0.13per million tokensopenrouter-models, last checked 2026-09-24 for input against $0.18per million tokensopenrouter-models, last checked 2026-09-24 on the preview, which is still listed beside it — the generally available row listing below the preview it replaced — and Wikipedia records that Hy3 was made available under the Apache License in July 2026, following an April 2026 previewsource, accessed 2026-08-28. Both are the top listed provider's rate for their row rather than necessarily Tencent's own, so this is a gap between two listings and not between two prices Tencent set. The preview is the expensive way to use the same generation, and it stays on the shelf after the cheap one arrives.

In the catalog as observed on 31 August 2026, tencent/hy4-preview was the only one of Tencent's seven rows with no Hugging Face id. The other six all had one, tencent/hy3-preview included. The empty field was the catalog lagging, not the weights being held back: Tencent open-sourced the model the day it announced it, and instruct weights and an FP8 variant published on Hugging Face, ModelScope, GitCode and CNBsource, accessed 2026-09-02. The first snapshot this site holds in which the row carries tencent/Hy4-preview was fetched on 8 September. So Tencent shipped downloadable weights at preview stage again. It is also the only Tencent row listing above a dollar per million output tokens: $2.50per million tokensopenrouter-models, last checked 2026-09-24 against $0.53per million tokensopenrouter-models, last checked 2026-09-24 for the generally available third generation. Published weights let third parties host the row as they host Hy3: on 25 September 2026 OpenRouter's endpoint list named Novita and SiliconFlow beside Tencent, all three at the same price. The oldest row in the set, tencent/hunyuan-a13b-instruct from 8 July 2025, is still listed and still carries its weights.