Tencent Hy4 Flash Weights Are Out: What Hardware Runs It?

A compact computer glowing cyan on a wooden desk in front of a larger dark server tower, with a Tencent Hy logo badge in the corner

Tencent published the Hy4 flash weights on Hugging Face on 17 November 2026: tencent/Hy4-flash, tencent/Hy4-flash-FP8 and tencent/Hy4-flash-base, none of them gated (tencent/Hy4-flash, 2026). The model card lists 220B parameters in total with 11.3B active per token for text, and a 1M-token context. The official 4-bit GGUF from Tencent's AngelSlim team, Q4_K_M, is 156.79 GiB, which fits a 512GB Mac Studio or two DGX Spark with room for context, but not a 128 GB machine. One catch: llama.cpp support is still an open pull request, and mlx-lm has none yet.

Update, 18 November 2026: this article was first published on 17 November, when the weights were announced but not yet on Hugging Face. It now uses the real file sizes and the official model card.

What is in the Hy4 flash release?

From the model card of tencent/Hy4-flash:

SpecValue
Total parameters220B
Main weights (excluding Engram, ViT, DFly2)155.5B
Engram parameters63.5B
Vision encoder (ViT)0.43B
DFly2 drafter0.85B
Active per token11.3B text + 0.43B vision
Experts per MoE layer512 routed (16 active) + 1 shared
Context1M tokens (1,048,576)
InputText and image

The figures differ slightly by source. The model card text says 220B total, including 64B Engram, and 12B active per token. Its spec table gives 11.3B active for text plus 0.43B for vision, about 11.7B, which rounds to 12B. The total is 220B: the Hugging Face API counts 220,251,201,958 parameters in the published weights (tencent/Hy4-flash, 2026), and the four parts in the table add up to the same, 155.5 + 63.5 + 0.43 + 0.85 = 220.28B. Tencent's press release gives 224B, a figure the weights do not match. Tencent describes the model as natively multimodal: it takes text and images in and writes text out. Engram is a table of learned bigram and trigram embeddings, looked up by hash, that sits next to the MoE backbone. The DFly2 drafter is a small built-in model for speculative decoding that reuses the main model's KV cache. Tencent says the hybrid attention design cuts the KV cache to 1/24 of Hy3's and 1/6 of Hy4's, but publishes no figure in gigabytes.

Three official repos are up, with sizes read from the Hugging Face API on 18 November:

RepoFormatSize
tencent/Hy4-flashBF16 safetensors, 148 shardsabout 441 GB
tencent/Hy4-flash-FP8FP8 safetensorsabout 287 GB
tencent/Hy4-flash-baseBF16 base modelabout 439 GB

In exact terms the BF16 repo holds 440.52 GB of safetensors and the FP8 repo 287.07 GB.

What license is Hy4 flash under?

The Tencent Hy License. It is derived from Apache 2.0 but is not Apache 2.0: the license text says so itself, and adds commercial terms in Section 10. If you run a Model-as-a-Service business (remote access to inference or fine-tuning, for example through an API), you need a separate written license from Tencent once your worldwide revenue passes US$50 million over 12 months, or before using the model in a product with more than 100 million monthly active users. Running it on your own machine is not a MaaS business. Read the LICENSE before you build a hosted service on it.

How big are the GGUF builds?

The AngelSlim/Hy4-flash-GGUF repo, from Tencent's AngelSlim compression team, has five builds. Sizes are from its README, in GiB:

BuildSizeIn GB
Q4_K_M (recommended default)156.79 GiBabout 168 GB
SPARK-4Bit151.3 GiBabout 162 GB
SPARK-4Bit-Fast148.6 GiBabout 160 GB
Q3_K144.70 GiBabout 155 GB
IQ2-AH125.73 GiBabout 135 GB

AngelSlim calls Q4_K_M "the safe default" and the SPARK-4Bit builds "suited for Spark devices". The Q4_K_M comes as two files, 104.99 GB and about 63 GB. The second file has the same size, to the byte, in the Q3_K and IQ2-AH builds too, which matches the 63.5B-parameter Engram table at about 8 bits per weight: only the main weights change between the builds (our reading of the file list).

Which hardware can hold Hy4 flash at Q4?

A GiB is 1.0737 GB, so the Q4_K_M build is 156.79 × 1.0737 ≈ 168.4 GB of weights, before macOS or Linux, apps and the KV cache.

  • Two DGX Spark (2 × 128 GB = 256 GB): 256 − 168.4 ≈ 88 GB left for the system and context. It fits.
  • 512GB Mac Studio: 512 − 168.4 ≈ 344 GB left. It fits with a wide margin, enough for long context.
  • 256GB Mac Studio: the same 88 GB margin as two DGX Spark, on one machine.
  • 128 GB Mac: 168.4 − 128 ≈ 40 GB short before the operating system even loads. Q4 does not fit. Even the smallest build, IQ2-AH at about 135 GB, is larger than 128 GB.

All margins are our calculation from the file sizes. The KV cache grows with the context you load, and Tencent gives no gigabyte figure for it, so we do not estimate one.

What speeds does AngelSlim report?

AngelSlim's README gives these numbers. They are AngelSlim's own tests, not ModelFit measurements, and we have not reproduced them.

Machine and buildDecode tok/s (4K / 16K / 32K context)
Mac M5 Max 128GB, Q4_K_M48.13 / 45.18 / 44.15
Mac M5 Max 128GB, Q3_K50.34 / 49.47 / 47.37
RTX Spark 128GB, SPARK-4Bit-Fast31.95 / 33.14 / 33.39
RTX Spark 128GB, SPARK-4Bit26.28 / 30.23 / 35.38

Prompt processing on the M5 Max is 629.8 / 448.1 / 418.6 tok/s for Q4_K_M at the same three context lengths, and 799 / 717 / 707 tok/s for SPARK-4Bit-Fast on the RTX Spark.

Read these with care. Both test machines have 128 GB, and every build they ran is larger than 128 GB, which is the memory math above. AngelSlim's README does not say how the runs were set up. Our reading is that the roughly 63 GB Engram file, a lookup table, is read from SSD rather than held in memory: without it, the Q4_K_M main weights are 104.99 GB. AngelSlim does not confirm this, so we do not count a 128 GB Mac as a supported machine. If you want the whole model in memory, plan for 256 GB or more.

Will llama.cpp, MLX, vLLM and SGLang run it?

EngineStatus on 18 November 2026
llama.cppNot merged: pull request #32517 is open
mlx-lmNo support
vLLMOfficial image vllm/vllm-openai:hy4-flash
SGLangOfficial image lmsysorg/sglang:hy4-flash

llama.cpp: the upstream pull request #32517, "Model : add hy_v4_flash support", is open and not merged. AngelSlim's README tells you to build from that PR branch. The GGUF repo also ships two patches, hy4-flash-mac-minimal.patch for the Mac and hy4-flash-spark.patch for Spark. A stock llama.cpp release will not load these files yet.

mlx-lm: there is no Hy4 flash model in mlx_lm/models and no MLX build from Tencent. The older hunyuan.py there is a different architecture.

vLLM and SGLang: Tencent recommends them for production serving and provides official Docker images. These target data center GPUs, not Macs.

So the only Mac route today is a patched or PR-branch llama.cpp. ModelFit will add Hy4 flash to its model catalog once #32517 is merged or mlx-lm adds support.

How does it compare with the Hy4 preview?

Hy4 previewHy4 flash
Total parameters770B220B
Active per token49B11.3B (text)
4-bit GGUF416.97 GB (IQ4_XS)about 168 GB (Q4_K_M)
LicenseApache-2.0Tencent Hy License
llama.cppMerged (hy_v4)PR #32517, open
Smallest machine at 4-bit512GB Mac Studio256GB Mac Studio or 2 DGX Spark

The preview needs a 512GB Mac Studio at 4-bit, as we covered in Tencent Hy4 Open Weights: Can a 512GB Mac Studio Run It?. Hy4 flash takes about 40% of that memory and activates under a quarter of the parameters per token, which matters for speed because every token reads the active weights from memory.

FAQ

Are the Hy4 flash weights on Hugging Face?

Yes. Tencent published tencent/Hy4-flash, tencent/Hy4-flash-FP8 and tencent/Hy4-flash-base on 17 November 2026, not gated, under the Tencent Hy License.

How much memory does Hy4 flash need?

The official Q4_K_M GGUF is 156.79 GiB, about 168 GB, before the system and the KV cache. The smallest build, IQ2-AH, is 125.73 GiB, about 135 GB.

Can a Mac run Hy4 flash?

A 256GB or 512GB Mac Studio holds the Q4_K_M build in memory. A 128 GB Mac is about 40 GB short at Q4. On any Mac, you need llama.cpp built from pull request #32517 or with AngelSlim's patch, because support is not merged yet.

Is Hy4 flash Apache 2.0?

No. It uses the Tencent Hy License, which is derived from Apache 2.0 and adds commercial terms for large Model-as-a-Service businesses in Section 10.

How much does the Hy4 flash API cost?

About US$0.070 per million input tokens and US$0.281 per million output tokens, per Tencent.

Sources

  • tencent/Hy4-flash model card, file list and LICENSE: https://huggingface.co/tencent/Hy4-flash
  • Tencent-Hunyuan/Hy4-flash on GitHub: https://github.com/Tencent-Hunyuan/Hy4-flash
  • tencent/Hy4-flash-FP8: https://huggingface.co/tencent/Hy4-flash-FP8
  • tencent/Hy4-flash-base: https://huggingface.co/tencent/Hy4-flash-base
  • AngelSlim/Hy4-flash-GGUF README and file list: https://huggingface.co/AngelSlim/Hy4-flash-GGUF
  • llama.cpp pull request #32517: https://github.com/ggml-org/llama.cpp/pull/32517
  • mlx-lm models directory: https://github.com/ml-explore/mlx-lm/tree/main/mlx_lm/models
  • Tencent, Tencent Launches Hy4 flash: https://www.tencent.com/tencent-launches-hy4-flash-to-power-high-speed-cost-efficient-ai-deployment-for-everyday-workflows/
  • tencent/Hy4-preview model card: https://huggingface.co/tencent/Hy4-preview
What hardware runs this?

Match this model to a machine that can run it: by RAM tier for Apple Silicon, or by VRAM for an NVIDIA GPU.

See how this changes your recommendation

Run the wizard

Have questions? Reach out on X/Twitter