Hy4's launch headline reads "joining the ranks of the strongest open-source models" — but nine days after launch (Nov 13), the full-release weights still can't be downloaded. This is the easiest point of the Hy4 launch to misunderstand — and the reason several aggregators list the preview weights as the full release's. This article sets out the current state, the evidence and the confusions, one by one.
On Nov 5, digitalapplied found tencent/Hy4 returning 401 on Hugging Face, with no Hy4 repository on GitHub; on Nov 7, OrcaRouter wrote "the version number went up, the openness came down" — the preview open-sourced on launch day, while the full release ships API-first, weights later; on Nov 11, EveryInfra re-checked Hugging Face and still found only preview weights. All three independent checks agree: the full-release weights were unpublished as of Nov 13.
Tencent's official wording is "coming soon" ("in the coming days" in English), with a system card to follow the weights. Partners include vLLM, SGLang and Inferact Inc. — which suggests inference-framework support will likely be ready when the weights ship. But "soon" has no date, no system card has appeared, and the full-release license is likewise unpublished. Artificial Analysis' model page still labels Hy4 as proprietary.
Hy4 preview was released and open-sourced on August 28, 2026, with weights on Hugging Face (tencent/Hy4-preview, plus an FP8 variant), ModelScope, GitCode and CNB under Apache 2.0. BF16 is about 1.56TB, FP8 about 814GB (EveryInfra's figure, Nov 11; another aggregator writes about 924GB — this page uses 813.8GB and notes the source). Running FP8 needs roughly 500GB-class VRAM hardware.
But the preview's limitation is critical: it is a text-only model, with no vision understanding and no computer use — the two headline capabilities of this full release. Local deployment, self-hosting or fine-tuning with the preview is possible; running the full release's vision and computer use locally is not. The full release's weight format, quantization variants and hardware requirements can only be confirmed once the weights ship. One aggregator claims a self-hosting floor of 16× B200 or 8× B300, but that has a single source and is unverified — this page does not present it as fact.
"Hy4 is already open source." Not true. The full-release weights are unpublished; only the preview is downloadable. Several aggregators (byteiota, verticaltabs) list the preview weights as the full release's — that's the mix-up.
"The 770B parameters are officially announced for the full release." Partly true. The Nov 4 release post did not restate the full release's parameter count or layer architecture; 770B/49B comes from Tencent Cloud's techpedia (phrased as the "Hy4 series") and the preview model card. The figure is likely reliable, but strictly speaking the system card is needed for confirmation.
"Hy4 is open-sourced under Apache 2.0." Partly true. Apache 2.0 is the preview's license; the full-release license is unpublished. Don't transfer the preview's license to the full release.
"Hy4's weights are on Hugging Face." Partly true. What's on Hugging Face is tencent/Hy4-preview, not tencent/Hy4 — the latter returns 401.
Tencent has given no explanation for the delay. Reasonable guesses: the full release adds vision and computer use, and weight formats and inference-framework adaptation take time; writing the system card and completing safety evaluations take time; and after the day-one usage limits on Nov 5, Tencent prioritized expanding the inference cluster (tenfold). But these are guesses — Tencent has offered no explanation or timetable. Notably, Kimi K3.1 Pro's weights shipped the same day it launched, a contrast that makes Hy4's delay look sharper.
When the weights do ship, four things are worth checking first: first, whether the license is Apache 2.0 or something more restrictive; second, the weight format and quantization variants (BF16 / FP8 / NVFP4?) and hardware requirements; third, the system card's confirmation of parameter count and layer architecture — whether 770B/49B applies to the full release; fourth, independent local benchmark reproductions — the official 44.4% Terminal-Bench 4.0 and AA's 37% will get a third number from local reproduction. This page will be updated when the weights ship.
Reference: digitalapplied — Tencent Hy4 Is Out: What Is Confirmed and What Is Not Yet (rewritten and organized by ModelBrief)