Newsletter image

Subscribe to the Newsletter

Join 10k+ people to get notified about new posts, news and tips.

Do not worry we don't spam!

By pressing the Subscribe button, you confirm that you have read and are agreeing to our Privacy Policy and Terms of Use

Search

GDPR Compliance

We use cookies to ensure you get the best experience on our website. By continuing to use our site, you accept our use of cookies, Privacy Policy, and Terms of Service.

Tencent - Multi-Modal, AI Coding, Ecosystem

Tencent Hy4

Tencent's Hy4 is 770B parameters with 49B active and a 1M context. We diffed the licence against Apache 2.0: identical through Section 9, with one MaaS clause added. And Hy4-flash has a 63.4 GB Engram floor that no quantisation touches.

TL;DR
  • 770B total, 49B active per token, 1M context, image and video input
  • Licence is Apache 2.0 verbatim through Section 9, plus one MaaS clause
  • Hy4-flash: 220B total, 12B active, and a 63.4 GB floor quantisation cannot move
System Requirements
RAM135GB for the smallest complete Hy4-flash GGUF, 63.4GB of it fixed
GPU8-GPU node for Hy4-flash-FP8; the 770B flagship is rented, not local
VRAM230GB

Tencent pushed the Hy4 weights to Hugging Face at 03:05 UTC this morning: 770B total parameters, 49B active per token, image and video understanding, a 1M token context window, and 1.5 TB of safetensors that almost nobody reading this will download. The size is the least interesting thing about the release. The interesting thing is a plain text file called LICENSE, which we diffed against Apache 2.0 line by line. It is Apache 2.0, verbatim, with exactly one section added. For the Hy3 generation seven months ago, the equivalent file did not grant you any rights at all if you were in the European Union.

What shipped, and when

This is a family release spread over three days, not a single drop. Hy4 flash came first and has had most of the attention; the flagship landed this morning. Both ship in bf16 and FP8, both ship a pre-trained base model alongside the instruct model, and both are multimodal.

Repository What it is Params (total / active) Weights on disk Weights published Likes
tencent/Hy4 Flagship instruct 774B / 49B 1.55 TB 2026-11-20 03:05 UTC 71
tencent/Hy4-FP8 Flagship, FP8 774B / 49B FP8 2026-11-20 03:05 UTC 9
tencent/Hy4-base Flagship, pre-trained 774B / 49B bf16 2026-11-20 03:05 UTC 8
tencent/Hy4-flash Cost-tuned instruct 220B / 12B 440 GB 2026-11-19 11:25 UTC 392
tencent/Hy4-flash-FP8 Cost-tuned, FP8 220B / 12B FP8 2026-11-19 11:28 UTC 22
tencent/Hy4-flash-base Cost-tuned, pre-trained 220B / 12B bf16 2026-11-18 04:24 UTC 24

The download counters tell you where the real traffic is going. The official tencent/Hy4-flash repository had 2,053 downloads when we checked. The GGUF conversion at AngelSlim/Hy4-flash-GGUF had 55,872, which is 27 times more. AngelSlim is Tencent's own model-compression group rather than a third party racing ahead of the vendor, so this is not a packaging gap. It simply says that essentially everyone touching this model is going through llama.cpp rather than the safetensors.

There is also a GitHub repository at Tencent-Hunyuan/Hy4, created on 2026-10-28 and last pushed at 02:46 UTC this morning, 19 minutes before the weights went up. It holds Python code, 4.4 MB of it, and had 54 stars and zero open issues when we looked. Issues are enabled, which is a change from several recent Chinese lab releases we have covered.

The licence is Apache 2.0 with exactly one section bolted on

The file is titled "Tencent Hy License, Version 1.0" and its own preamble says it is derived from Apache 2.0 "with modifications, mainly the additional commercial terms for large Model-as-a-Service ("MaaS") businesses in Section 10", and that it is not the Apache License and not endorsed by the Apache Software Foundation. That is an unusually honest preamble, and we checked whether it is accurate.

We pulled the canonical Apache 2.0 text from apache.org, extracted the terms and conditions from both files, normalised whitespace, and diffed them. The result is short enough to state completely. Sections 1 through 9 are Apache 2.0 word for word, with four changes:

  • Section 1 says "Sections 1 through 10" where Apache says "1 through 9".
  • The definition of "Work" gains a sentence putting it beyond doubt that the Work includes "the model weights, parameters, configuration files, inference and training code, and documentation".
  • The definition of "Derivative Works" gains a sentence stating that these include "any fine-tuned, quantized, distilled, merged, pruned, or otherwise modified versions of the model weights".
  • Section 4, on redistribution, gains an anti-circumvention clause: any licence terms you put on the Work or a derivative "must not remove, limit, or conflict with Section 10, and Section 10 shall continue to apply to every recipient".

That is the entire diff outside Section 10. The first three changes are clarifications that help you rather than restrict you, because they settle the question of whether weights are copyrightable subject matter under this grant in the direction of "treat them as covered". The fourth matters: Section 10 is sticky. It travels with every quantisation, every fine-tune, every merge, to every downstream recipient, and you cannot relicense it away.

So what is Section 10? It applies only if you run a MaaS business, defined as giving third parties remote access to model inference or fine-tuning, for example via an API, where they submit their own inputs, parameters or training data. The definition explicitly excludes two things: end-user products where model capabilities are "confined to specific features or workflows, without general-purpose model access", and merely relaying requests to somebody else's model service.

If you do run a general-purpose MaaS business, you need a separate written licence from Tencent in two cases: your aggregate worldwide gross revenue exceeds US$50 million in any consecutive 12-month period, whether or not any of it comes from the model; or your product has more than 100 million monthly active users in any calendar month. Section 10.3 exempts internal use entirely, provided neither the model nor its outputs reach third parties.

Read that back against what you are probably building. A SaaS product with an AI feature is not MaaS. A coding assistant is not MaaS. Running the model inside your company is explicitly exempt. Fine-tuning it and shipping the fine-tune is governed by Sections 1 to 9, which is to say Apache 2.0. The threshold for the one restriction is a revenue figure that excludes virtually every independent developer and most funded startups. For practical purposes this is Apache 2.0 with a clause aimed at hyperscalers.

What the previous licence said, and who it shut out

This is where the change registers, because the comparison is not against Apache 2.0 but against what Tencent shipped before. We fetched the LICENSE from tencent/Hy3-preview, the direct predecessor flagship, released 2026-04-23. Its third line is in capitals:

THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED
KINGDOM AND SOUTH KOREA AND IS EXPRESSLY LIMITED TO THE TERRITORY,
AS DEFINED BELOW.

"Territory" is then defined as the worldwide territory excluding the EU, the UK and South Korea, and Section 5(c) adds that you "must not use, reproduce, modify, distribute, or display" the model, its outputs or its results outside that Territory, and that any such use "is unlicensed and unauthorized". The same text still governs HunyuanVideo-1.5, which is where we ran into it before.

Term Hy3 preview (2026-04-23) Hy4 (2026-11-20)
Base Bespoke community agreement Apache 2.0, verbatim, plus Section 10
Territory Excludes EU, UK, South Korea No territorial limit
Acceptable Use Policy Incorporated by reference, Tencent may update it at any time None
Downstream obligations Must pass use restrictions on as enforceable provisions Apache notice requirements, plus Section 10 flow-down
Trademark Must use the "Tencent Hy" mark as required by Section 3(c) Standard Apache Section 6, no grant, no requirement
Large-user gate Over 100M MAU: request a licence, granted at Tencent's sole discretion Over 100M MAU or US$50M revenue, and only for general-purpose MaaS
Grant Non-exclusive, non-transferable, limited Perpetual, worldwide, irrevocable, sublicensable

The intermediate step is worth recording because it stops this being a simple story. Hy4 preview, published 2026-08-27, was plain Apache 2.0 with a Tencent preamble and no territorial language at all; we checked that file too and it contains zero occurrences of "Territory", "European Union" or "Acceptable Use". So Tencent dropped the territorial exclusion in August, not today. Measured against Hy4 preview, today's licence is slightly tighter, because Section 10 did not exist in August. Measured against the Hy3 generation, it is a different universe: a perpetual irrevocable sublicensable grant with no geography, no acceptable use policy and no discretionary approval, where there used to be a limited non-transferable one that excluded Europe outright.

Our own export-ban piece carries a table row listing "Tencent Hy3" under "Hunyuan Community (not OSI)". That row was correct when we wrote it and is now out of date for the current generation. The Hunyuan media models still use the old agreement; the language models do not.

It is still not open source, and that matters less than usual

Section 10 is a field-of-use restriction, so this licence would not pass the Open Source Definition and Tencent does not claim it would. Call it open weights, not open source, the way we do for MiniMax M3.1 Flash.

The difference is in where the line sits. MiniMax M3.1 Flash is non-commercial by default with a US$20 million gate on commercial use of any kind, which catches real companies doing ordinary things. Hy4's gate is US$50 million of revenue and operating a general-purpose inference API. Those two conditions almost never co-occur outside firms with a legal department already on this. If you are choosing a model for a product and the licence is your constraint, Hy4 is now closer to DeepSeek's MIT releases than to the custom-licence pack it used to sit in.

Hy4: 770B total, 49B active, and 21 of 78 layers that actually look

The language backbone is 78 layers: one dense feed-forward layer, then 77 MoE layers with 256 routed experts and one shared expert each, activating the top 8 routed experts plus the shared one per token. Hidden size 6,144, vocabulary 120,832, context 1,048,576 tokens. Attention is Multi-head Latent Attention with q_lora_rank 2,048 and kv_lora_rank 512, wrapped in what Tencent calls Gated DeepSeek Sparse Attention, with a 32-head indexer selecting the top 2,048 tokens of context. Residual flow uses identity Hyper-Connections with four parallel residual streams.

The sparse-attention lineage is explicit: the card cites the DeepSeek-V3.2 paper for DSA and a second paper, IndexCache, for cross-layer index reuse. That second one is the part worth pulling out, because the config file tells you how aggressive the reuse is in a way the card does not. The indexer_types array has 78 entries, one per layer, each either full or shared. Counting them: 21 layers compute a fresh sparse index and 57 reuse one. The full layers sit at indices 0, 1, then every fourth layer from 5 to 77. So for roughly three quarters of the network, deciding which 2,048 tokens to attend to is a cache lookup rather than a computation. That is where the long-context economics come from.

On top of the backbone sit a 1.23B parameter ViT with dynamic resolution support and a 3.3B parameter five-layer drafter called DFly2 for speculative decoding. DFly2 is the neat piece of engineering here: each draft layer shares both the KV cache and the indexer's top-2,048 selection with its corresponding backbone layer, so drafting costs a constant amount regardless of context length and adds no KV memory of its own.

Hy4-flash is the one you might actually run

Flash is not a distillation of the flagship with the same shape. It is a different architecture, and three of the differences are interesting.

Property Hy4 Hy4 flash
Backbone / total / active 770B / 774B / 49B text 155.5B / 220B / 11.3B text
Layers 78 (1 dense + 77 MoE) 42 (1 dense + 41 MoE)
Hidden size 6,144 3,072
Routed experts, top-k 256, top-8 512, top-16, in a 1,536-dim latent space
Attention layout Gated DSA everywhere, 21 of 78 indexing 12 sparse + 30 sliding-window, 512-token window
N-gram embedding None Engram, 63.4B params at layers 1, 11, 21, 31
Vision encoder 1.23B, 40 layers, trained in-house 0.43B, 27 layers, adapted from SigLIP2-So400m
Drafter DFly2, 3.3B DFly2, 0.85B
Context 1M 1M

The LatentMoE idea, running routed experts in a lower-dimensional latent space, is the same family of trick as the rebuilt expert layer in Edge0-145B, arrived at independently. The hybrid attention layout interleaves sparse and sliding-window layers at a 1:3 ratio, in a stated pattern of two sparse layers followed by ten repeats of three sliding-window layers and one sparse layer. The payoff Tencent claims is the number to remember: KV cache reduced to 1/24 of Hy3 and 1/6 of Hy4. At a 1M context that is the difference between a model you can serve and one you cannot.

Engram is the genuinely novel piece. The backbone is augmented with learned bigram and trigram representations retrieved through multi-head hash lookups: eight hash heads per n-gram order, a 10.3 million entry hash vocabulary per head, 768 dimensions per order, injected at four layers. It is 63.4B parameters of pure embedding table, 29% of the model's total parameter count, and it does no matrix multiplication at all. Which leads to the thing we found by accident.

The 63.4 GB that quantisation cannot touch

The GGUF conversions ship in two shards. Looking at the file sizes in the AngelSlim repository, the second shard is 63,360,616,128 bytes in the IQ2_AH build, in the Q3_K build, and in the Q4_K_M build. Identical to the byte, across three different quantisation levels. Hugging Face exposes the LFS SHA256 for each file, so this is checkable rather than inferred, and all three resolve to the same hash:

ecd6a0ecd7f258b1ee9211bf566261ae33efc1743ad40f58cb5004cf7b051839   Hy4-flash-IQ2_AH-00002-of-00002.gguf
ecd6a0ecd7f258b1ee9211bf566261ae33efc1743ad40f58cb5004cf7b051839   Hy4-flash-Q3_K-00002-of-00002.gguf
ecd6a0ecd7f258b1ee9211bf566261ae33efc1743ad40f58cb5004cf7b051839   Hy4-flash-Q4_K_M-00002-of-00002.gguf

It is the same file. Shard two is the Engram table, shipped once and reused: 63,360,616,128 bytes over 63.4B parameters works out to 0.9994 bytes per parameter, which is 8 bits, matching the engram_fp8_embedding: true flag in the config. Quantisation only ever touches shard one.

Build Shard 1 (quantised) Shard 2 (Engram, fixed) Total download Engram share
Hy4-flash IQ2_AH 71.6 GB 63.4 GB 135.0 GB 46.9%
Hy4-flash Q3_K 92.0 GB 63.4 GB 155.4 GB 40.8%
Hy4-flash Q4_K_M 105.0 GB 63.4 GB 168.3 GB 37.6%
Hy4-flash SPARK-4Bit single file 162.5 GB n/a
Hy4 STQ1_0 single file 230.5 GB n/a
Hy4 IQ2 single file 332.4 GB n/a
Hy4 Q4_K_M single file 470.2 GB n/a

Two consequences. First, "220B total, 12B active" sounds like a model that shrinks nicely, and it does not: there is a hard floor at 63.4 GB that no quantisation level moves, so the smallest honest Hy4-flash download is 135 GB and nearly half of it is a lookup table. Going from Q4_K_M to IQ2_AH saves 33 GB, not the 60% a parameter count would suggest.

Second, and more useful: because Engram is hash-indexed embedding lookup rather than weights in a matmul, it is the ideal candidate for staying off the accelerator. Keep the 63.4 GB in system RAM or memory-mapped from an NVMe and put only the 71.6 GB of quantised backbone in VRAM, and the hardware story changes shape. We have not measured what that does to tokens per second, and the access pattern is random, so an NVMe will hurt more than host RAM will. But the split is architecturally clean in a way that a dense 135 GB model's is not. If you have 96 GB of VRAM and 128 GB of system memory, this is the model to try that trick on.

Running it

Both models have day-one vLLM and SGLang recipes, with prebuilt images, which is a better launch posture than most. Note the tensor-parallel-size of 8 in both official examples: these are eight-GPU commands.

docker run --gpus all -p 8000:8000 --ipc=host \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  vllm/vllm-openai:hy4-flash tencent/Hy4-flash-FP8 \
    --tensor-parallel-size 8 \
    --speculative-config '{"num_speculative_tokens":5,"method":"dfly"}' \
    --attention-backend FLASHMLA_SPARSE \
    --tool-call-parser hy_v4 \
    --reasoning-parser hy_v4 \
    --enable-auto-tool-choice \
    --port 8000 \
    --served-model-name hy4-flash

For the GGUF path, pull both shards and the vision projector. llama.cpp picks up shard two automatically when you point it at shard one, but the mmproj file is separate and without it you have a text-only model.

hf download AngelSlim/Hy4-flash-GGUF \
  Hy4-flash-IQ2_AH-00001-of-00002.gguf \
  Hy4-flash-IQ2_AH-00002-of-00002.gguf \
  mmproj-Hy4-flash-bf16.gguf \
  --local-dir ~/models/hy4-flash

# confirm shard two is the Engram table everyone else also has
sha256sum ~/models/hy4-flash/Hy4-flash-IQ2_AH-00002-of-00002.gguf

llama-server -m ~/models/hy4-flash/Hy4-flash-IQ2_AH-00001-of-00002.gguf \
  --mmproj ~/models/hy4-flash/mmproj-Hy4-flash-bf16.gguf \
  -c 32768 --host 127.0.0.1 --port 8080

Once a server is up, either path speaks the OpenAI API. The one Hy4-specific control is reasoning effort, which defaults to deep chain-of-thought and has to be turned off explicitly rather than turned on:

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="hy4-flash",
    messages=[{"role": "user", "content": "Summarise this changelog in three bullets."}],
    temperature=0.9,
    top_p=1.0,
    extra_body={"chat_template_kwargs": {"reasoning_effort": "no_think"}},
)
print(response.choices[0].message.content)

Tencent's recommended sampling is temperature=0.9 and top_p=1.0, which is looser than most instruct models want. reasoning_effort takes "high" or "no_think" and nothing in between.

What the numbers say, and what the appendix hides

Every figure below is Tencent's own. There are no third-party evaluations of either model yet, which is expected one day after release, and worth keeping in mind for longer than that.

The pre-training comparison in the card is the most useful table because it includes the predecessor and both DeepSeek V4 base models. Hy4 Base leads or ties 24 of the 27 benchmark rows, losing only OCW Courses, MBPP-plus and Belebele. Where it does not lead is the more informative part:

Benchmark Hy4 Base Hy4 flash Base Hy3 preview Base DeepSeek-V4-Pro-Base
MMLU-Pro (5-shot) 73.9 71.1 65.8 73.4
SuperGPQA (5-shot) 62.7 57.8 52.3 54.8
SimpleQA (5-shot) 57.2 37.5 26.4 52.1
MBPP-plus (3-shot) 74.0 73.7 77.4 79.1
OCW Courses (5-shot) 66.7 64.0 62.6 68.2
MMMU-Pro (4-shot) 59.6 52.0 not multimodal not multimodal

MBPP-plus is the row to stare at. Hy4 Base scores 74.0 against 77.4 for its own predecessor and 79.1 for DeepSeek-V4-Pro-Base, so on this particular code benchmark the new flagship base model is the worst of the four. SimpleQA going from 26.4 to 57.2 generation over generation is the largest single jump in the table and suggests a lot more world knowledge in pre-training. Tencent prints its own caveat in bold under the table: the base model saw "a large amount of synthetic instruction data during mid-training", so it sometimes breaks the expected answer format in few-shot evaluation, which may distort the scores. Credit for saying so, but it does mean this table is softer than it looks.

The other thing to flag is the appendix. Both cards have a section headed "Appendix: benchmark results", and in both cases the content is a JPEG. The full post-training benchmark tables, the comparison models and the evaluation notes are all pixels. You cannot grep them, diff them against a later release, or check an individual figure without reading numbers off an image. Several labs do this now and it is worth naming each time: a benchmark table published as a picture is a benchmark table that cannot be audited.

On the internal evaluations, the honest summary is that flash is cheap rather than close. On WorkIndex, Tencent's own eight-set internal benchmark covering slide creation, document editing, spreadsheet repair and visual understanding, Hy4 flash scores 42.1 against 39.7 for Hy4 preview and 51.4 for Hy4. The release note says flash lands "slightly above Hy4 preview", which is true, but the flagship is 18.1% higher than flash on the same index. What flash buys you is a cost per task of 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 at Max tier peak-hour pricing. In a blind evaluation with 85 internal users over 259 tasks, flash scored 2.79 out of 4 on satisfaction against 2.81 for Claude Haiku 5.5, at a median 6.7 minutes per task and RMB 0.26 per task.

The finance evaluation is the most unusual thing in the flagship card and deserves a mention because of how it was run: 579 real tasks, assessed by 213 finance professionals including 153 from financial institutions, more than 70% with over five years of experience. Hy4 averaged 3.04 out of 4, ahead of DeepSeek-V4.1-Flash-1017 at 2.97 and both Kimi K3.1 Pro and GLM-5.3 at 2.68. Tencent also reports cutting its own search hallucination rate from 15.76% to 5.45% between Hy4 preview and Hy4. Those are internal numbers with no published methodology, but a panel of that size is more effort than most vendor evaluations involve.

Limitations and gotchas

  • The flagship is not a local model under any quantisation. 1.55 TB in bf16, and the most aggressive community GGUF is still 230.5 GB. The eight-GPU tensor-parallel examples in the card are the realistic floor.
  • Flash has a 63.4 GB floor. No quantisation level touches the Engram table, so 135 GB is the smallest complete download and the parameter count will mislead you about memory.
  • Section 10 is sticky. It flows to every recipient of any derivative and you cannot licence it away. If you redistribute a fine-tune, you have to carry it.
  • AngelSlim's own GGUF repositories declare no licence at all in their Hugging Face metadata, while a third-party repack we checked correctly carries tencent-hy-license. Section 4 requires Section 10 to reach every recipient, so the missing metadata on Tencent's own compression repos is a gap on their side, not permission on yours.
  • Reasoning is on by default at reasoning_effort="high". If you are benchmarking latency or token cost without passing no_think, you are measuring deep chain-of-thought.
  • The official site was unreachable from our network. Both hy.tencent.ai/research/hy4 and the Hy4 system card PDF returned an HTTP 567 on every attempt, so we could not read the system card. The Hugging Face cards, the config files and the licence are all first-party and all reachable, which is what everything above is built from.
  • No independent benchmarks exist yet. Every performance figure in this article is Tencent's, including the ones comparing Tencent to its competitors.

Who should run which

If you are serving a product on rented GPUs: Hy4-flash-FP8 on an eight-GPU node, through the official vLLM image, with DFly2 speculative decoding enabled. This is the configuration the cost claims are based on, and the 1/6 KV cache against the flagship is what makes long contexts affordable.

If you have a workstation with serious memory: Hy4-flash IQ2_AH, and try putting the Engram shard in host RAM. 96 GB of VRAM plus 128 GB of system memory is the interesting configuration to test, and nobody has published numbers for it yet.

If you need the ceiling: Hy4, rented, for long-horizon coding and agent work. Tencent's own guidance is to use flash for everyday productivity tasks and the flagship for complex software engineering, and the 51.4 against 42.1 on WorkIndex supports that split.

If the licence was your blocker: re-read it. If you were avoiding Hunyuan language models because of the territorial exclusion or the acceptable use policy, both are gone, and have been since August. The current terms are Apache 2.0 unless you are running a general-purpose inference API with over US$50 million in revenue.

Sources and further reading

The ten-minute version: open the LICENSE file before the model card. Four paragraphs of Section 10 decide whether this 770B model is usable in your product, and for almost everyone reading this the answer is yes in a way it was not seven months ago. Then, if you are going to run flash locally, check the size of shard two before you plan your hardware.

Not independently tested. Every figure here comes from the Tencent model cards, the config and licence files in the weights repositories, the Hugging Face and GitHub APIs, the canonical Apache 2.0 text, or our own arithmetic and diffs on those. No part of either model was run on our hardware, and no third-party benchmarks of Hy4 or Hy4-flash existed at the time of writing.
Date checked: 2026-11-20

Prev Article
Haru-4B
Next Article
OpenThinker-32B

Related to this topic: