A generated title card reading 'Hy4 flash is open weight', with three repository tiles labelled 'Hy4 flash', 'Hy4 flash-FP8' and 'Hy4 flash-Base', a spec strip reading '220B total / 12B active / 1M context / image in, text out', and a footer reading 'Weights published 17 November 2026 under the Tencent Hy License v1.0.'
Guides & Insights

Hy4 flash Is Open Weight Now: Read Section 10 Before You Download It

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

The Tencent Hy License, Version 1.0, is an Apache 2.0 derivative with one extra section, and that section is the reason to read this release carefully rather than just pulling the checkpoint. Section 10.2 says that if you operate a model-as-a-service business and your worldwide gross revenue exceeds US$50 million in any twelve-month period — or your product passes 100 million monthly active users — you need a separate written licence from Tencent before you use the weights, or continue using them. Section 10.3 exempts internal use. Section 2 of the Apache grant is where the royalty-free rights live, and Section 10.2 opens with the words "Notwithstanding Sections 2 and 3." So the download is free and the deployment is not unconditionally free, and the line between the two is drawn at whether strangers can reach the model through your infrastructure.

That is the frame for what Tencent published on 17 November 2026 at 03:44 UTC, when three repositories appeared on Hugging Face within ninety seconds of each other: Hy4 flash, the instruction-tuned chat model; Hy4 flash-FP8, the same model in FP8; and Hy4 flash-Base, the pre-trained checkpoint with no post-training at all. The instruction model is 220 billion total parameters with 64 billion of them in Engram tables, 12 billion active per token, natively multimodal on the input side, and configured for a 1,048,576-token context. It is the same model Tencent put on its API two days earlier, on 16 November, with the difference that you can now run it on hardware you own.

What is actually in the three repositories

The split is cleaner than most open-weight releases, because the three checkpoints are genuinely different things rather than three filenames for one artefact.

• Hy4 flash — the instruct model, chat template and reasoning modes included. Supports reasoning_effort at high (the default, thinking on) and no_think (direct answers). Tencent recommends temperature 0.9 and top_p 1.0.

• Hy4 flash-FP8 — the same weights quantised for production serving. This is the one the official vLLM and SGLang recipes point at, and the one Tencent says clears 150 tokens per second on a tensor-parallel-8 node.

• Hy4 flash-Base — pre-trained only. The card is explicit that this "is a pre-trained base model, not an instruction-tuned chat model," which matters because every benchmark table in the release is a base-versus-base table. The instruct model's numbers are the vendor's WorkIndex and blind-test figures, not a public eval suite.

Two details in the metadata are worth more than they look. The first is that the repositories are flagged image-text-to-text at the pipeline level, so the vision encoder ships with the weights rather than being bolted on downstream. The second is a numbering discrepancy: Tencent's launch material on 16 November described the model as 224B, and the model card's own table sums to 220.3B — 155.5B backbone, 63.5B Engram, 0.43B ViT, 0.85B DFly2 drafter. Neither number is wrong in context; the 224B figure appears to count the DFly2 drafter and rounding differently, while the card's table excludes Engram, ViT and DFly2 from its own 155.5B backbone line and then adds them back. When a vendor ships two totals for one model, the card is the one to quote.

A generated reference card headed 'What the three repositories actually contain', listing the three checkpoints with a one-line description each and a separate outlined panel headed 'Tencent Hy License v1.0, section 10.2' noting the separate written licence above US$50M annual gross revenue or above 100M monthly active users as a service.

The architecture is a cost argument wearing an architecture's clothes

Very little of this design makes sense read as a capability play. Read as a way to shrink the KV cache and the decode step, all of it does.

• Attention — sparse IndexPool attention interleaved with 512-token sliding-window attention at a 1:3 ratio, over a multi-head latent attention (MLA) formulation. Sparse layers select up to 2,048 tokens by group-level selection; there are 12 sparse layers and 30 SWA layers across the 42-layer backbone.

• Routing — 512 routed experts plus one shared expert per MoE layer, 16 routed activated per token, operating in a LatentMoE space of dimension 1,536 rather than the full expert width.

• Engram — learned bigram and trigram representations retrieved by multi-head hash lookups at backbone layers 1, 11, 21 and 31. That is the 63.5B the card counts separately from the backbone.

• Drafter — DFly2, a 5-layer 0.85B speculative-decoding model that reuses the target model's sparse KV cache and indexer selection rather than keeping its own. Tencent's claim is a constant draft cost of about 2,048 keys per draft token, regardless of how long the context is, and zero extra KV memory.

• Headline efficiency figure — KV cache at 1/24 of Hy3's and 1/6 of Hy4's, vendor-reported and not independently measured.

• Post-training — self-distillation from Hy4 plus reinforcement learning, which Tencent says cuts average end-to-end output tokens by 50%. Shorter outputs are where a large share of the cost saving actually comes from, and it is the part least visible in any per-token price card.

Put together, that is a model engineered so that a 1M-token context does not cost what a 1M-token context normally costs to serve. The context length is the marketing line; the KV cache fraction is the engineering claim underneath it.

Screenshot of the Hugging Face model card for tencent/Hy4-flash, showing the three-model release line, the Model Links table and the quickstart snippet.

What Tencent measured, and what it did not

Every number in this section is Tencent's, produced on Tencent's harness, and none of it has been reproduced outside the company. Artificial Analysis has no page for Hy4 flash — the model URL returns 404 as of 18 November — so there is no third-party score to set beside any of it.

The base-model table compares Hy4 flash-Base against Hy4 Base, Hy3 preview Base, DeepSeek-V4-flash-Base and DeepSeek-V4-Pro-Base, all on one pipeline. It is the most honest table in the release precisely because it loses rows: MMLU-Pro reads 71.1 for Hy4 flash-Base against 73.9 for Hy4 Base and 73.4 for DeepSeek-V4-Pro-Base; SimpleQA reads 37.5 against 57.2 and 52.1. Where it wins, it wins by a lot on Chinese-language and multilingual rows — C-Eval 92.0, CMMLU 91.6, AGIEval Chinese 90.6, Belebele 93.8, INCLUDE 82.8 — which is what you would expect from a model built for Tencent's own products first.

Read the small print on the card before you read the scores. Tencent states that heavy exposure to synthetic instruction data during mid-training leaves the model occasionally unable to follow "the expected answer format in few-shot evaluations, which may distort the scores" — a vendor flagging that some of its own rows may be measuring format compliance instead of capability. Treat every number below as an upper bound until your own harness agrees.

The instruct model has no public benchmark table at all. What it has instead is WorkIndex, an internal suite of eight task families — slide creation, slide editing, human-AI slide co-writing, Markdown co-writing, document editing, spreadsheet repair, audio/video processing, and visual understanding. Hy4 flash scores 42.1 there against 39.7 for Hy4 preview and 51.4 for Hy4 itself, at an average cost per task the company puts at 1/13 of Hy4 preview, 1/20 of Hy4 and 1/8 of DeepSeek-V4.1-Flash-1017 at peak-hour pricing. And there is a blind evaluation with 85 internal WorkBuddy users across 259 tasks in 27 categories, where Hy4 flash scores 2.79 out of 4 for satisfaction against 2.81 for Claude Haiku 5.5, at a median 6.7 minutes and RMB 0.26 per task.

Those are real measurements of a real population. They are also measurements of Tencent employees doing Tencent-shaped work, scored on a rubric Tencent wrote, and the 0.02-point gap to a competitor's satisfaction score is far inside any plausible noise band. Treat the WorkIndex ordering as a directional signal from the vendor and nothing more.

The licence is where this release will actually be judged

Open weights are not open source, and the Tencent Hy License makes the distinction unusually concrete. Four clauses decide what a given team can do:

• Base grant — Apache 2.0's permissions, including commercial use, modification and redistribution, apply to the weights, parameters, configuration, inference code and documentation.

• The MaaS trigger — a separate written licence is required if you provide third-party remote access to inference or fine-tuning and exceed US$50 million in worldwide gross revenue in any twelve-month window. Revenue derived from the model counts, and so does revenue that has nothing to do with it: the clause reads "whether or not such revenue is derived from the Work."

• The scale trigger — the same requirement applies before using the weights in any product with more than 100 million monthly active users in a calendar month, at any revenue level.

• Two carve-outs — end-user products whose model access is confined to specific features rather than general-purpose prompting are outside the MaaS definition, and "merely relaying requests to an independently operated third-party model service" is excluded too. Internal-only use is exempt outright.

Compare that with the licence on the ERNIE 5.5 Flash Base checkpoint, which excludes the EU, the UK and South Korea as territories entirely. Tencent does not exclude any territory. It taxes scale instead. For a startup building a product on top of the weights, that is a materially friendlier document than a territorial ban. For a mid-sized inference provider, it is a phone call to [email protected] before shipping.

Serving it: three real paths, all of them pinned

The deployment story is more complete than the launch-week API story was, and it is also more fragile than a supported product, in the specific way that early framework support always is.

• vLLM — the official image vllm/vllm-openai:hy4-flash, tensor-parallel 8, the FP8 checkpoint, --attention-backend FLASHMLA_SPARSE, DFly speculative decoding at 5 tokens, and the hy_v4 tool-call and reasoning parsers.

• SGLang — lmsysorg/sglang:hy4-flash, multi-arch across x86 and Arm, tensor-parallel 8, --speculative-algorithm DFLY with block size 6.

• Transformers — not yet in a release. The card tells you to install from a pull-request ref (refs/pull/50903/head) until the merge lands, which is the clearest possible signal of how new this is.

• Quantisation — a GGUF build from the AngelSlim toolkit is linked for end-to-end testing, and a fine-tuning guide ships in the repository.

Pinned images and a PR ref mean the model works today on the exact builds Tencent tested and may not work on the next minor release of either framework. That is worth knowing before you put it behind a production endpoint, and it is why the framework recipes matter more than the benchmark table for anyone actually deploying.

What the open weights do to the API price

Tencent's mainland list price for Hy4 flash is ¥0.5 per million input tokens, ¥2 per million output, and ¥0.02 for a cache hit — roughly $0.070, $0.281 and $0.003 at current rates. Those are the vendor's card, not a measurement.

That price is not what changes when weights ship. What changes is the floor under it. A model you can download has a second price, which is the cost of the GPUs you run it on, and the entire architecture described above is an argument that this second price is unusually low for a 1M-context model. Tencent's own 1/8-to-1/20 per-task cost ratios are that argument in vendor arithmetic.

OrcaRouter routes 200+ models behind one API with the provider's list price passed through at 0% markup, so a vendor price change is live on our side the same day rather than at the next catalogue rebuild. Hy4 flash itself is not on that catalogue yet — what is there is tencent/hy4-preview, the 770B/49B teacher model this one was distilled from, at $0.834 input and $2.501 output per million with a $0.042 cache-read rate. If you want to evaluate the family before committing to a self-hosted FP8 deployment, the preview is the model you can call today; the weights are what you move to when the numbers hold up. Automatic failover is the other half of that posture, because a two-day-old checkpoint running on a PR-ref Transformers build is exactly the kind of dependency you want a second route behind.

Screenshot of the tencent/hy4-preview model page, showing the per-million-token input, output and cache-read prices the routed preview model is listed at.

What to watch

Three things will decide whether this release matters in six months, and none of them are the benchmark table.

First, the 404. Hy4 flash has no independent score. Until Artificial Analysis or an equivalent runs it, every capability claim in this piece is Tencent's, and the honest position is that the model is unmeasured rather than measured-well. Second, the framework merges: when the Transformers support leaves pull request 50903 and the vLLM and SGLang images stop being pinned, the model stops being a research artefact and becomes infrastructure. Third, the licence interpretations — the first company to run a $60 million MaaS business on these weights will tell everyone how Section 10.2 is actually read.

Tencent closed its announcement with the line "stay tuned for Hy4 coming up next," which is a promise that the 51.4-on-WorkIndex model in the cost charts gets the same treatment. If it does, the interesting question stops being whether Tencent can ship a cheap open-weight flash model and becomes whether it can ship an expensive one that competes on capability rather than price.

Compared in this article2

Detected from this article · Benchmarks: Artificial Analysis · updated daily