Hy4 flash open weights: what a million-token context does not solve
Tencent's Hy4 flash ships open weights, image input and a one-million-token context. Three things enterprises should check before putting documents into the prompt: how the model selects context, what the licence means for products, and what the vendor benchmarks do not measure.
CogLake Redaktion · 22 November 2026 · 6 minThis article was created automatically with AI support.
The headline number is a context window. The number that decides whether it works in your company is smaller. Tencent has published the weights of Hy4 flash — the instruction-tuned model, an FP8 build and a base model for further training — on Hugging Face, ModelScope and GitCode. TechNode reported the release on 19 November 2026 and dates it to Wednesday, 18 November, two days after the model went live in Tencent's own apps and cloud API. The model card lists 220 billion total parameters, of which about 11.3 billion are active per token for text, plus a context window of 1,048,576 tokens and image input.
For enterprise knowledge work that sounds like the end of retrieval: put the folder in the prompt and skip the index. The model card says why that is the wrong conclusion.
What was actually published
| Artefact | What it is for |
|---|---|
| Hy4 flash | Instruction-tuned model, text and image input, 1M-token context |
| Hy4 flash-FP8 | Quantised build for serving and for end-to-end testing |
| Hy4 flash-base | Pre-trained base model, not a chat model — the starting point for your own post-training |
| Tencent Hy License | Licence text in the repository; Apache 2.0 derived, with additional commercial terms |
Also on the card: official recipes for vLLM and SGLang, an OpenAI-compatible serving path, a GGUF build for testing, and a finetuning guide. What the card does not contain is a minimum hardware requirement.
A million tokens, but not a million tokens of attention
The architecture is the interesting part for anyone who runs retrieval. Hy4 flash interleaves sparse attention with sliding-window attention at a 1:3 ratio. The sliding-window layers see a 512-token causal window; the sparse layers select up to 2,048 positions through a group-level indexer. Every token therefore does not attend to everything in the window. Selection happens inside the model, per token, over what it has been given.
That design is what makes the long window affordable: the card states the KV cache is reduced to 1/24 of Hy3 and 1/6 of Hy4, and that the bundled speculative-decoding drafter attends to at most about 2,000 keys per draft token regardless of context length. These are vendor figures without a measurement protocol, but the direction is plausible — sparse attention is how every long-context release of the past year has bought its window.
The enterprise consequence is not technical trivia. If you put a whole document set into the prompt, the decision about which passages matter is made by a learned indexer, not by your access rules, your document hierarchy or your retention policy. Retrieval you control gives you a citable passage, a permission check and a log line. Attention you do not control gives you a plausible answer. For questions that end up in an audit, that is a downgrade, not an upgrade.
Where the cost numbers come from
The card's efficiency claims rest on WorkIndex, an internal benchmark built from Tencent's own productivity tasks — slide creation and editing, document editing, spreadsheet repair, audio/video processing. On that set Hy4 flash scores 42.1 against 39.7 for Hy4 preview and 51.4 for Hy4, at an average cost per task of 1/13 of Hy4 preview, 1/20 of Hy4 and 1/8 of DeepSeek-V4.1-Flash-1017 at peak-hour pricing. A blind evaluation with 85 internal users over 259 tasks in 27 categories puts user satisfaction at 2.79 out of 4, against 2.81 for Claude Haiku 5.5, at a median of 6.7 minutes and RMB 0.26 per task.
None of this is a price per token, and none of it transfers automatically to German or English contract work. The honest reading: this is a model positioned for volume, and the vendor's own recommendation follows — Hy4 flash for everyday productivity and deep search, the larger Hy4 for long-horizon software engineering.
The licence is the second product decision
The Tencent Hy License is derived from Apache 2.0 and adds commercial terms in Section 10. Two paragraphs matter:
- §10.2 requires a separate written licence for a model-as-a-service business if worldwide gross revenue exceeded US$50 million in any consecutive 12-month period, or if a product has more than 100 million monthly active users. §10.2 applies whether or not the revenue comes from the model.
- §10.3 exempts internal use, provided neither the model, its derivatives nor their outputs are made available to third parties.
"Derivative Works" explicitly includes fine-tuned, quantised, distilled, merged and pruned versions of the weights. A team that quantises the model for its own cluster and later offers the resulting endpoint to customers is therefore in scope, not outside it. Internal document search, contract review and knowledge assistants that only serve employees are the case §10.3 covers.
That is a licensing statement, not a compliance statement. Data protection, works council consultation, retention and the processor relationship are separate questions that a licence text does not answer.
What the benchmarks do not show
The base-model table is the most useful part of the card, because it shows what the "flash" trade costs. On knowledge benchmarks the smaller model sits below Hy4 Base: MMLU 88.7 against 91.6, MMLU-Pro 71.1 against 73.9, MMMU-Pro 52.0 against 59.6 — and on SimpleQA, which measures short factual answers, 37.5 against 57.2.
Tencent also prints a caveat under the table: the base model saw a large amount of synthetic instruction data during mid-training and occasionally fails to follow the expected answer format in few-shot evaluation, which may distort scores. A vendor warning about its own numbers is worth more attention than the numbers.
The practical reading is unchanged from every other model release: a long window and a low price per task do not raise factual precision. If your use case is a claim that has to be correct, the check has to live outside the model.
Five questions before a pilot
| Question | Why it decides the pilot |
|---|---|
| Do you need the long window, or do you need a citation? | Sparse selection is not evidence. If answers must be traceable, keep retrieval and use the window only for reading a retrieved set. |
| Which passages did the model see? | Log the prompt, not just the answer. With 1M tokens the prompt itself becomes an audit artefact. |
| Who is the licensor for your scenario? | Internal use is covered by §10.3; a customer-facing endpoint above the thresholds is not. |
| What is your hardware baseline? | The card gives no minimum. The official recipes assume eight-way tensor parallelism on the FP8 build — treat that as the starting assumption, then measure. |
| What is the fallback model? | An open-weight model you cannot serve at your volume is a dependency with a licence, not an exit option. |
What this article does not prove
We did not run Hy4 flash. Everything here comes from the model card, the licence file and TechNode's report; the efficiency figures are Tencent's own, measured on Tencent's tasks, with no independent replication. Whether the model is good at German-language contract text is an open question — the card lists German only as part of a five-language evaluation subset, not as a primary target.
And there is a cheaper test than a pilot: take twenty questions your team already answered with sources, run them through the model once with the whole document set and once with your existing retrieval, and compare how often the cited passage actually supports the answer. That comparison decides more than any benchmark table.
This article is based on published vendor documentation and licence terms; we did not run our own tests. It is not legal advice — licence and data protection questions belong with your counsel.
Sources
- Tencent, model card "tencent/Hy4-flash", Hugging Face, including architecture table, benchmark tables and licence reference: https://huggingface.co/tencent/Hy4-flash
- Tencent, "Tencent Hy License, Version 1.0", §§10.2–10.4 (MaaS, internal use): https://huggingface.co/tencent/Hy4-flash/raw/main/LICENSE
- TechNode, "Tencent releases Hy4 flash model weights for local deployment", 19 November 2026: https://technode.com/2026/11/19/tencent-releases-hy4-flash-model-weights-for-local-deployment/
Sources retrieved on 22 November 2026.