Sulat.com / Writing on AITuesday, November 17, 2026
Artificial Intelligence/16 November 2026

Tencent launches Hy4 flash at US$0.070 per million input tokens, with open weights promised

Tencent's Hy team launched Hy4 flash, a 224B-parameter model with 12B active per token and a 1M-token context, at US$0.070 per million input tokens. The weights aren't out yet.

Glossy blue 3D Tencent Hy logo sphere above gold coins and a price tag, ringed by glass slide, document and spreadsheet icons

At a glance

  • Tencent's Hy team launched Hy4 flash on 16 November, a multimodal model with 224 billion parameters, 12 billion active per token and a one-million-token context window, available through Tencent Cloud TokenHub and Tencent's apps.
  • Hy4 flash costs US$0.070 per million input tokens and US$0.281 per million output tokens, against US$0.834 and US$2.501 for Hy4. Cached input costs US$0.003, Tencent says.
  • In a blind test with 85 internal WorkBuddy users, Tencent says Hy4 flash scored 2.79 on a four-point satisfaction scale against Claude Haiku 5.5's 2.81, at 0.26 yuan a task against 4.73.
  • DeepSeek-V4.1-Flash-1017 scored highest on satisfaction in Tencent's test, at 3.09, and took longest, with a median of 10.3 minutes. Hy4 flash had the shortest median time, at 6.7 minutes.
  • Tencent says it plans to release Hy4 flash as an open-weight model but hasn't given a date. Free WorkBuddy access and 50% off TokenHub API pricing end on 30 November.

Tencent’s Hy team has launched Hy4 flash, a multimodal model with 224 billion parameters, 12 billion of them active for each token, and a one-million-token context window, Tencent says in its launch post. The API costs US$0.070 per million input tokens and US$0.281 per million output tokens, against US$0.834 and US$2.501 for Hy4.

Tencent announced it on X on 16 November. The post pitches Hy4 flash as the cheapest model for everyday productivity and says it matches Claude Haiku 5.5 on daily work in human blind tests, at an 18x lower cost per task.

What’s inside

Hy4 flash is a mixture-of-experts model, so only part of it runs for each token: 12 billion of the 224 billion parameters. Tencent counts 64 billion of the total as Engram parameters. The architecture combines hybrid attention, LatentMoE and Engram, which Tencent says cuts the KV cache, the memory a model keeps for text it has already read, to 1/24 of Hy3’s size and 1/6 of Hy4’s.

In post-training, Tencent used self-distillation from Hy4, then reinforcement learning to cut average end-to-end output tokens by 50%, it says. For serving, Tencent lists megakernels, DFly and 4-bit quantisation, which stores a model’s weights in fewer bits. Together they push decoding past 150 tokens per second.

The blind test

Tencent’s headline claim comes from a blind evaluation with 85 internal WorkBuddy users across 259 tasks in 27 categories. Tencent says Hy4 flash’s user satisfaction came out close to Claude Haiku 5.5’s, with the shortest end-to-end completion time and a far lower cost per task than any other model tested. The external models ran at their high tier.

ModelSatisfaction (4-point)Median timeCost per task (yuan)
DeepSeek-V4.1-Flash-10173.0910.3 min1.20
Hy4 flash2.796.7 min0.26
Claude Haiku 5.52.816.8 min4.73
GLM-5.3-Flash2.537.6 min0.81
Blind-test results from Tencent’s X post. The three other models ran at their high tier.

Haiku 5.5 scored 2.81 against Hy4 flash’s 2.79 on the four-point scale. Haiku’s cost per task was 4.73 yuan against 0.26, which is where the 18x comes from. DeepSeek-V4.1-Flash-1017 scored highest on satisfaction, at 3.09, and took longest, with a median of 10.3 minutes.

Cost per task in Tencent's blind test: Haiku 5.5 4.73 yuan, DeepSeek 1.20, GLM 0.81, Hy4 flash 0.26, an 18x gap
Claude Haiku 5.5 cost 4.73 yuan a task in Tencent’s test; Hy4 flash cost 0.26.

WorkIndex and Tencent’s own teams

Tencent also measured Hy4 flash on WorkIndex, an internal benchmark built from typical WorkBuddy tasks. It has eight sets: slide creation, slide editing, human-AI slide co-writing, Markdown co-writing, document editing, spreadsheet repair, audio and video processing, and visual understanding. Tencent says Hy4 flash scores slightly above Hy4 preview overall, and its average cost per task is 1/13 of Hy4 preview’s, 1/20 of Hy4’s and 1/8 of DeepSeek-V4.1-Flash-1017’s, on that model’s Max tier at peak-hour pricing. The same figure charts seven public agent benchmarks, among them Terminal-Bench 4.0, HLE and OSWorld 2.1.

Tencent’s own product teams report results too. WorkBuddy says Hy4 flash cuts average task time by 70% at roughly 1/15 of Hy4’s cost, matching Hy4’s quality on everyday tasks with a slight dip on long-horizon coding. The WeCom assistant Dayuan saw output speed rise 88% over Hy4, with about 60% fewer output tokens in thinking mode. In WeChat Pay’s tests against DeepSeek-V4.1-Flash-1017, static code scanning finished 37% faster on 46% fewer tokens. All three are Tencent teams reporting on Tencent’s own tests.

Price and access

Hy4 flash is live on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, and Tencent says it’s coming to OpenRouter soon. Through 30 November, WorkBuddy users get it free and TokenHub takes 50% off API pricing.

Price per million tokens, Hy4 flash vs Hy4: input US$0.070 vs 0.834, output 0.281 vs 2.501, cached input 0.003 vs 0.042
Tencent lists Hy4 flash input at about a twelfth of Hy4’s price.

Tencent recommends Hy4 flash for deep search, document processing, slide creation, simple frontend demos and computer use. It still recommends Hy4 for complex work such as long-horizon software engineering, because of Hy4’s higher intelligence ceiling.

Hy4 flash for deep search, documents, slides, simple frontend demos and computer use; Hy4 for long-horizon software engineering
Tencent points everyday work at Hy4 flash and long software jobs at Hy4.

Open weights

Tencent says it plans to release Hy4 flash as an open-weight model and to ship frequent updates. It hasn’t given a date. Tencent’s Hugging Face account lists Hy4-preview and its FP8 version, but not Hy4 flash. Hy4 itself isn’t there either, though Tencent said when it launched Hy4 earlier this month that it planned to release those weights too.

Tencent says the OpenRouter listing is coming soon, and the free WorkBuddy and half-price TokenHub offers end on 30 November.


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.