Field guide Hy4 Flash

Checked

Unofficial guide

Hy4 Flash vs Hy4

Last updated

Hy4 Flash is the smaller, cheaper sibling of Hy4, not a rename of it. The numbers in the tables are from Tencent's model cards, the 16 November 2026 announcement, and the TokenHub pricing page. This guide did not run the benchmarks.

Architecture

Hy4 Flash figures are from the Hy4-flash model card. Hy4 figures are from the tencent/Hy4 model card. Hy4 preview's headline size is from Tencent's announcement and the Hy4-preview repository, which describes the preview backbone.
Hy4 Flash Hy4 Hy4 preview
Total parameters 220B on the card, including 63.5B Engram, 0.43B ViT, and 0.85B DFly2. Announcement says 224B. 770B language backbone, excluding ViT, adapter, and DFly2. ViT 1.23B. DFly2 3.3B. 770B total, 49B active, in the announcement and the preview README.
Active per token Text 11.3B, plus vision 0.43B. Headline figure is 12B. Text 49B, plus vision 1.23B. 49B in the announcement. The preview README's backbone table says 49B activated.
Layers 42: 1 dense and 41 MoE. 78: 1 dense and 77 MoE. 78 in the preview README.
Experts 512 routed, 16 active, plus 1 shared. Latent dimension 1,536. 256 routed, 8 active, plus 1 shared. 256 routed, 8 active, plus 1 shared, in the preview README.
Hidden size 3,072 6,144 6,144 in the preview README.
Attention Hybrid sparse IndexPool and sliding window, MLA. Gated DeepSeek Sparse Attention with IndexCache, MLA. Gated DSA in the preview README.
Context 1,048,576 1,048,576 1M
Modalities Text and image in, text out. Text, image, and video understanding, text out. The preview README reviewed here describes the language backbone. The later Hy4 card is the one that states video.
Drafter DFly2, 0.85B, 5 layers. DFly2, 3.3B, 5 layers. The preview README describes 1 native MTP layer, 10B total and 0.7B activated.

The Flash card lists three changes relative to Hy4: hybrid attention at a 1:3 sparse-to-sliding-window ratio, LatentMoE, and Engram n-gram embeddings. It also says the KV cache is 1/6 the size of Hy4 and 1/24 the size of Hy3.

API list price

USD per million tokens from the English TokenHub pricing page updated 19 November 2026. A 50 percent discount on Hy4 Flash was scheduled through 12:00 Beijing time on 30 November 2026, and on Hy4 through 31 December 2026.
Model id Input Output Cache
hy4-flash 0.07 0.281 0.003
hy4 0.834 2.501 0.042

Scores Tencent published

On WorkIndex, an internal benchmark of WorkBuddy-style tasks, the model card reports 42.1 for Hy4 Flash, 39.7 for Hy4 preview, and 51.4 for Hy4. It says Flash's average cost per task is 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 on that benchmark's Max tier at peak-hour pricing. The announcement repeats the preview's size, 770 billion total and 49 billion active, and the same cost ratios. WorkIndex covers slide creation and editing, document editing, spreadsheet repair, audio and video processing, and visual understanding. The card's “open benchmarks” group is DeepSWE, ProgramBench, Terminal-Bench 4.0, APEX-Agents, AutomationBench, HLE, and OSWorld 2.1. The card does not print a numeric score for that group in the text this guide could extract, so none is repeated here.

A blind test described on the card used 85 internal WorkBuddy users, 259 tasks, and 27 categories. Hy4 Flash scored 2.79 out of 4, against 2.81 out of 4 for Claude Haiku 5.5, with a median end-to-end time of 6.7 minutes and a cost of 0.26 RMB per task. That is an internal study. It is not an independent leaderboard.

The base-model few-shot table is also Tencent's, and it comes with their caveat: the base model saw a large amount of synthetic instruction data in mid-training and sometimes misses the expected few-shot format, which can distort scores. Macro-averaged accuracy is marked on several text rows. A sample of the rows they printed:

Base checkpoints, not the chat models. Scores are Tencent's. MMLU and MMLU-Pro are 5-shot and marked as macro-averaged. MBPP-plus is 3-shot.
Benchmark Hy4 flash base Hy4 base Hy3 preview base DeepSeek-V4-flash-Base DeepSeek-V4-Pro-Base
MMLU 88.7 91.6 87.5 89.7 91.1
MMLU-Pro 71.1 73.9 65.8 68.6 73.4
SimpleQA 37.5 57.2 26.4 31.3 52.1
MBPP-plus 73.7 74.0 77.4 76.8 79.1

The full table, including the multimodal rows they report only for Hy4 Flash base and Hy4 base, is on the model card. Instruct-model scores against those DeepSeek checkpoints are not in the text this guide verified, so they are not invented here.

Common questions

Should I use Hy4 Flash or full Hy4?

Tencent recommends Hy4 Flash for most everyday work, including deep search, documents, slides, simple frontend demos, and computer use. They recommend full Hy4 when the ceiling matters, and they name long-horizon software engineering as the example. On their internal WorkIndex, Hy4 scores 51.4 and Hy4 Flash scores 42.1. On the TokenHub list price checked 19 November 2026, Hy4 Flash is 0.07 USD per million input tokens and Hy4 is 0.834.

How does Hy4 Flash compare with Hy4 preview?

Tencent's announcement describes Hy4 preview as 770 billion total parameters and 49 billion active. The Hy4 Flash model card says Flash scores 42.1 on WorkIndex and Hy4 preview scores 39.7, and that Flash's average cost per task on that internal benchmark is 1/13 of Hy4 preview. Those are Tencent's measurements.

Is Hy4 Flash the same size as other open models it is compared with?

No public comparison in Tencent's card treats it as the same architecture as DeepSeek-V4. The card does publish a base-model few-shot table that includes DeepSeek-V4-flash-Base and DeepSeek-V4-Pro-Base, and a cost comparison against DeepSeek-V4.1-Flash-1017. This site does not add scores of its own.

Sources

  1. tencent/Hy4-flash on Hugging Face Model card, architecture table, deployment commands, and license pointer.
  2. tencent/Hy4 on Hugging Face Full Hy4 model card used for the comparison.
  3. Tencent-Hunyuan/Hy4-preview on GitHub
  4. Tencent, “Tencent Launches Hy4 flash” Announcement dated 16 November 2026.
  5. Tencent Cloud, TokenHub model pricing English list prices. Page updated 19 November 2026.