Tencent opens Hy4 flash weights for self-hosting and fine-tuning
Edited and reviewed by AI Ping Editorial Team
3-line summary
- Standard, FP8 and base weights are available to download.
- Deployment and training guides specify substantial hardware needs.
- Large general model services face additional licensing terms.
Tencent released Hy4 flash weights on November 18, adding private deployment to its cloud API offering. Standard, FP8 and pretrained-base artifacts are available for download and adaptation.
The Tencent Hy License requires a separate written agreement for model-service businesses above either threshold: more than $50 million in aggregate worldwide revenue over any consecutive 12 months, or more than 100 million monthly active users for the relevant product.
The standard and FP8 editions are instruction-tuned; the base is for subsequent training. The description specifies text and image input, text output and up to one million context tokens. Its conversational examples apply to the instruction-tuned editions. About 12 billion parameters are active per token, according to Tencent.
Operating the downloads involves substantial storage and compute. The dedicated vLLM recipe lists checkpoint downloads of 440.5GB for the default edition and 287.1GB for FP8. Those are file sizes. GPU-memory estimates are not measured minima.
The default vLLM configuration uses eight H200 GPUs. It offloads part of the model to host RAM, requiring system memory alongside GPU memory. Although the checkpoint advertises a much larger context limit, the example server starts at 8,192 tokens, four concurrent sequences and one image per request. Larger settings change the memory budget.
Compression team AngelSlim has also published five GGUF builds, ranging from 125.73 to 156.79GiB, including standard quantizations and variants for Spark hardware. It reports its own Mac and RTX Spark tests. Download size, device memory and measured throughput describe different conditions; the document supplies no single universal memory requirement. Most non-expert weights in the smallest build remain in BF16.
AngelSlim links GGUF execution to a llama.cpp change awaiting merger. Its image-input example also needs a separate visual-projection file, and it provides distinct Mac and Spark instructions alongside the compressed downloads.
Inference support is initially supplied through dedicated releases. SGLang’s guide covers the standard and FP8 versions on H200, B300 and GB300 hardware, and says support has yet to enter a released package. It directs users to a dedicated image or model-support source branch, separating deployment configurations from additional experimental combinations. The experimental area additionally exposes host-memory or storage caching and separate input-processing and output-generation instances. These options sit outside the guide’s signed-off deployment combinations.
Tencent’s fine-tuning guide accepts text-only samples alongside image and video samples in a shared training file. Visual placeholders must match the linked files in order.
Training has a separate hardware requirement. For both full tuning and LoRA adaptation, Tencent lists a tested minimum of four machines with 32 GPUs, at least 96GB of memory per GPU and 2TB of host memory per machine. Longer sequences, higher-resolution images and more video frames increase resource needs.
Separate terms for general model services
The license defines model services as remote inference or fine-tuning where outsiders supply inputs, parameters or training data. It excludes feature-limited end-user products and simple relays to independently operated third-party services. Purely internal use is exempt from the additional size-based agreement requirement when neither weights, derivatives nor outputs reach third parties.
Sources
- Hy4 flash open-weight release announcement — November 18, 2026 — Tencent Hy
- Hy4 flash model description and release artifacts — Tencent Hy
- Tencent Hy License, Version 1.0 — Tencent Hy
- Hy4-flash serving recipe — deployment and memory conditions — vLLM
- Hy4-flash GGUF builds and testing notes — AngelSlim
- Hy4 flash fine-tuning guide — Tencent Hy
- Hy4-flash deployment cookbook — SGLang
