Subscribe
5 min read

Tencent's Hy4 Flash Weights Come With a $50 Million Clause

The Hunyuan family's first Hy4-generation open weights landed on Hugging Face on November 18 under a new Tencent Hy License that is Apache 2.0 until you cross 50 million dollars in revenue or 100 million monthly users.

Tencent's Hy4 Flash Weights Come With a $50 Million Clause

Tencent has open-sourced the weights of its newest workhorse model, and for the first time the Hunyuan family ships with a custom license attached. The Hy4 flash weights landed on Hugging Face, ModelScope and GitCode on November 18, covering an instruct model, an FP8 quantized variant and a pretrained base. They are released under the "Tencent Hy License," a document that is explicitly derived from Apache 2.0 but adds a commercial clause: companies that resell model access at scale must negotiate separately with Tencent.

The distinction matters because Tencent's previous big release did not carry it. Hy4 preview, the 770-billion-parameter mixture-of-experts model the company open-sourced in late August, went out under plain Apache 2.0. Hy4 flash is the first model of the current generation whose weights the public can download, and the license is the test case for how Tencent intends to keep "open" from subsidizing its largest competitors.

What the model is

Hy4 flash launched as an API model on November 16, billed at $0.070 per million input tokens, $0.281 per million output and $0.003 for cached input, roughly one-twelfth the flagship Hy4's input rate of $0.834 and about one-ninth its output rate of $2.501. The model is natively multimodal, accepts text and images, and supports a context window of 1,048,576 tokens.

Tencent describes it as a 224-billion-parameter model with 12 billion active per token. The model card breaks the number down more precisely: a 155.5B language backbone, a 63.5B "Engram" module, a 0.43B vision encoder adapted from SigLIP2, and a 0.85B speculative-decoding drafter called DFly2. Active parameters are 11.3B for text plus the vision encoder when images are in the prompt.

The Engram module is the most unusual piece. It is a lookup table of bigram and trigram embeddings (8 hash heads per order, 10.3M entries per head) injected at four of the 42 backbone layers. Most of the 64B parameters sit in tables that are looked up rather than computed on, which is how a model activates only about 11B parameters per token while carrying 220B-plus of total weights.

Flat illustration: a long wall of small navy squares with two orange columns running through a central arched gate

The rest of the architecture is built to make long-context serving cheap. The backbone interleaves 12 sparse attention layers, which select up to 2,048 relevant positions via a compressed indexer, with 30 sliding-window layers capped at 512 tokens, all in a multi-head latent attention formulation. Tencent says the layout cuts the KV cache to one-sixth of Hy4's and one twenty-fourth of Hy3's. Post-training distills from the flagship Hy4 and adds reinforcement learning that, per Tencent, halves average end-to-end output tokens. Inference uses fused megakernels, 4-bit quantization and the DFly2 drafter, which reuses the target model's KV cache and sparse indices, to push decoding past 150 tokens per second.

What the claims show

Tencent evaluated the model on WorkIndex, an internal benchmark of eight task families covering slide creation and editing, document and spreadsheet work, audio/video processing and visual understanding. Hy4 flash scores 42.1 against 39.7 for the Hy4 preview and 51.4 for full Hy4, at an average cost per task that Tencent puts at one-thirteenth of the preview's, one-twentieth of Hy4's and one-eighth of DeepSeek-V4.1-Flash-1017 at its Max tier's peak-hour price. A blind evaluation with 85 internal WorkBuddy users across 259 tasks gave Hy4 flash a satisfaction score of 2.79 out of 4 against 2.81 for Anthropic's Claude Haiku 5.5, with a median completion time of 6.7 minutes and a cost of 0.26 RMB per task.

Every one of those numbers is vendor-reported, and two of the three comparison points are the company's own products. The published base-model evaluations are more candid about the gap to the flagship: on few-shot MMLU the Hy4 flash base scores 88.7 against 91.6 for the Hy4 base, and on SimpleQA it drops to 37.5 against 57.2. Tencent itself recommends Hy4 flash for "deep search, document processing, slide creation, simple frontend demos, and computer use" while steering long-horizon software engineering to the flagship. A disclaimer on the model card notes that extensive synthetic data in mid-training can occasionally break answer-format compliance in few-shot tests, which may distort the published scores.

What the license says

The Tencent Hy License is Apache 2.0 with a tenth section bolted on. Fine-tuning, quantization, merging, pruning and distillation are all permitted, and derivatives may be distributed freely, with one string attached: the commercial clause travels with the weights, and downstream licenses "must not remove, limit, or conflict with Section 10."

Section 10 defines a "MaaS business" as giving third parties remote access to model inference or fine-tuning, for example via API. Two carve-outs keep it from swallowing ordinary products: end-user applications whose model access is confined to specific features or workflows, and services that merely relay requests to an independently operated third-party model. On top of that, purely internal use is exempt as long as neither the model nor its outputs reach third parties.

A separate written license from Tencent is required in two cases.

Flat illustration: a small figure stamping a wax seal onto a giant open ledger with a padlock floating above it

The first: any operator with aggregate worldwide gross revenue above $50 million in any consecutive 12-month period, whether or not a dollar of it comes from the model. The second: anyone who wants to run the weights in or for a product with more than 100 million monthly active users in any calendar month. The revenue test is entity-wide, not model-specific, which in practice gates hyperscalers, large SaaS vendors and most funded API platforms, while leaving the long tail of startups, researchers and on-premises deployments on something functionally equivalent to Apache 2.0.

The thresholds are not new thinking. Meta's Llama community license has used a 700-million-MAU trigger since 2024, and conditional open licenses have spread across the frontier labs since. What Tencent is doing is applying the pattern to a model it is simultaneously selling at aggressive list prices, with a free WorkBuddy promotion running through November 30 and 50% off API pricing on Tencent Cloud. The open weights double as the marketing floor for the paid API.

The release is also incomplete. The flagship Hy4's weights remain "planned," according to the November 4 announcement, which promised a comprehensive system card alongside them, and the Hy4 blog post from November 16 still listed flash as a future open-weight release two days before it shipped. Tooling arrived with the weights: prebuilt vLLM and SGLang images, a complete fine-tuning pipeline, and community GGUF builds, though Transformers support was still pending a pull request as of this week.

On the API side, the positioning is straightforward: a workhorse model priced at a fraction of the flagship's rates, with its strongest claims in everyday office productivity rather than frontier reasoning. The interesting part for anyone building on open weights is the license. Tencent kept the Apache grant, added a fine-tuning welcome mat, and drew a bright line at the point where open weights could undercut its own cloud business. Whether other Chinese labs copy the clause will say more about where "open source AI" is headed than any of the benchmark tables.

Sources