Tencent Open-Sources Hy4, a 770B Multimodal Flagship With Base Weights
Tencent open-sourced Hy4, Hy4-FP8 and Hy4 base: 770B MoE, 49B active, 1M context, text, images and video, under Tencent Hy License rather than Apache 2.0.
Tencent's Hy team has posted weights for its flagship Hy4 model on Hugging Face, ModelScope and GitCode, releasing the instruction-tuned Hy4, an FP8 build called Hy4-FP8, and the pretrained Hy4 base in one drop. The language backbone is a 770B mixture-of-experts model that activates 49B parameters per token, plus a 1.23B vision encoder and a 3.3B DFly2 drafter. It handles text, images and video, with a 1 million-token context window. The license is the Tencent Hy License, not Apache 2.0. The earliest post we found on X that mentioned these weights appeared at 10:55 Beijing time on November 20, and the Hugging Face cards were updated around 11:05. Tencent's official accounts did not publish a matching original post in this window.
This is not a model that appeared for the first time today, and it is not the same artifact as Hy4 flash or the earlier Hy4 preview. The Hy4 API was already live. When Tencent open-weighted Hy4 flash on November 18, it told people to stay tuned for Hy4. Preview was an earlier, text-oriented release under Apache 2.0. Flash is the everyday tier, with 220B total parameters and 12B active per token. Today's drop is the flagship itself. The 770B figure excludes the ViT, the adapter and DFly2. Hy4 base is a pretrained checkpoint, not an instruction-tuned chat model.
The base-model scores are Tencent's own few-shot table, and format failures can distort them
The model card illustrates pretraining quality with a bits-per-byte plot on an internal code and agent validation set. Lower is better, and models are arranged by activated parameter count. Read the official figure for the points; they are not transcribed here.
The same card publishes a few-shot table for Hy4 Base. On knowledge tests, MMLU is 91.6, MMLU-Pro 73.9, SuperGPQA 62.7, C-Eval 95.0 and SimpleQA 57.2, all above Hy4 flash Base, Hy3 preview Base, DeepSeek-V4-flash-Base and DeepSeek-V4-Pro-Base in that table. Multimodal numbers are reported only for Hy4 Base and Hy4 flash Base; MMMU-Pro (Standard, 10 options) is 59.6 versus 52.0. Hy4 Base does not lead every row: on OCW Courses, DeepSeek-V4-Pro-Base scores 68.2 against 66.7, and on MBPP-plus it scores 79.1 against Hy4 Base's 74.0, which also sits below Hy3 preview Base at 77.4. These figures come from Tencent's evaluation suite. Boldface marks the highest number in a row, not statistical significance. The card also warns that the base model saw a large amount of synthetic instruction data in mid-training, so it sometimes misses the expected answer format in few-shot tests, which can distort the scores.
Against preview, Tencent stresses vision and long-horizon work — and plots the gap to closed models
Compared with Hy4 preview, Tencent says it substantially scaled both pre-training and post-training, added vision, and improved instruction following, long-context understanding, long-horizon tasks and deep research. The main comparison figure places Hy4, preview and other models on the same set of evaluations; use that figure for the plotted scores.
A few of the write-up's numbers are specific. After work with Tencent Interactive Entertainment Group, Tencent says Hy4 improves 19.1% on GameDev and 9.8% on GameLogic versus preview, and mentions adaptation for Unreal Engine, Unity and Godot. On the finance side, a study run with Tencent Financial Technology (FiT) involved 213 finance professionals, 153 of them from leading financial institutions, and 579 real-world financial tasks. Hy4's average score is 3.04 out of 4, above DeepSeek-V4.1-Flash-1017 at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68, which Tencent cites as first among open-weight models. That is a vendor-organized expert study, not a public third-party leaderboard. On IRAB, an internal benchmark from RabyteTech, Tencent says Hy4 is 12.03 percentage points above preview. On its own tests, it reports the hallucination rate falling from 15.76% to 5.45%, and errors from guessing from 23.1% to 5.6%.
The limitations section is just as direct. Tencent calls Hy4 one of the strongest open-weight models available, then says a notable gap remains versus top proprietary models, especially on internal, high-difficulty sets for research-level math, long-horizon coding, model self-improvement and visual question answering. The figure above is Tencent's own comparison. The gains versus preview on the main chart are not the whole picture.
The license is no longer Apache; Tencent Hy License adds a MaaS threshold
The weights are released under Tencent Hy License 1.0. The text says it is derived from Apache 2.0, but is not the Apache License and is not endorsed by the Apache Software Foundation. Section 10 is the part that matters for products. If you operate a model-as-a-service business that gives third parties remote access to inference or fine-tuning, you need a separate written license from Tencent once your aggregate worldwide gross revenue exceeds US$50 million in any consecutive 12-month period, or before using the work in a product or service that has more than 100 million monthly active users in any calendar month. The revenue test is worldwide gross revenue, not revenue from this model alone. The clause does not apply to purely internal use, provided the model, derivatives and their outputs are not made available to third parties. Most in-house deployments and product features will not automatically cross that line, but this is not an unrestricted Apache grant. Large MaaS businesses have to negotiate separately.
For serving, Tencent documents vLLM and SGLang with an OpenAI-compatible API. The recommended sampling settings are temperature 0.9 and top_p 1.0. The reasoning mode reasoning_effort defaults to high; switch it to no_think for a direct answer. Those instructions apply to Hy4 and Hy4-FP8. If you want a pretrained starting point, use Hy4 base. Treating that checkpoint as a chat model out of the box is the wrong read.
What landed today is the flagship weight drop Tencent had already flagged: the same 770B multimodal model, now available as an instruction build, an FP8 build, and a base checkpoint. Pick the right line — Hy4, flash, or a preview you may still be running — and then check whether Section 10 of the Tencent Hy License actually applies to the business you are in.