The whole dayMon 16 Nov 2026
AI Compass
WeeklyArchiveAbout

Tencent launches 224B Hy4 flash multimodal MoE at $0.070 input

Live across Tencent Cloud TokenHub, WorkBuddy, Yuanbao and OpenRouter with a 1M context window, offering 50% promotional API discounts until 30 November 2026.

Glowing cobalt-and-ice-blue Tencent Hunyuan sphere hovering over a silicon wafer and compact KV-cache prism linked to camera, document and audio glass tiles
Image: AI Compass illustration (Tencent Hunyuan marks)

Tencent Hunyuan released Hy4 flash on 16 November 2026 across Tencent Cloud TokenHub, OpenRouter (tencent/hy4-flash), WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, following its flagship multimodal Hy4 launch. The natively multimodal Mixture-of-Experts checkpoint pairs 224 billion total parameters (including 64 billion Engram parameters) with 12 billion active parameters per token across a 1-million-token context window.

LatentMoE KV-cache compression, 150 tok/s serving and API pricing

Pre-training combines hybrid attention, LatentMoE and Engram lookups to shrink KV-cache memory to one-sixth of Hy4 and one-twenty-fourth of Hy3, while self-distillation from Hy4 and reinforcement learning trim average end-to-end output tokens by 50%. At inference, megakernels, 4-bit quantisation and DFly push decoding throughput above 150 tokens per second.

Horizontal bar chart showing Hy4 flash costing 0.2 RMB per task with a 42.1 WorkIndex score, 13.8 times cheaper than Hy4 preview at 3.2 RMB (39.7) and 24.5 times cheaper than Hy4 at 5.7 RMB (51.4)
On WorkIndex, Hy4 flash scores 42.1 above Hy4 preview (39.7) at 0.2 RMB per task, 24.5 times cheaper than full Hy4 (5.7 RMB).Image: Tencent

Standard API rates per million tokens are $0.070 uncached input, $0.281 output and $0.003 cached input, down from $0.834 input, $2.501 output and $0.042 cached on Hy4; through 30 November 2026, Tencent Cloud TokenHub and OpenRouter bill 50% off ($0.035 input, $0.1405 output and $0.0015 cache read), and WorkBuddy usage is free. On Tencent’s eight-set WorkIndex benchmark, Hy4 flash edges past Hy4 preview at one-twentieth the per-task cost of Hy4 and one-eighth that of DeepSeek-V4.1-Flash-1017 (Max peak pricing). Tencent stated that it plans to release Hy4 flash as an open-weight model at a later date.

Table of 259-task WorkBuddy human blind test results showing Hy4 flash scoring 2.79 satisfaction in 6.7 minutes at 0.26 RMB per task versus Claude Haiku 5.5 at 2.81 in 6.8 minutes and 4.73 RMB
Across 259 blind tasks with 85 WorkBuddy users, Hy4 flash matched Claude Haiku 5.5 satisfaction (2.79 vs 2.81) at 0.26 vs 4.73 RMB a task.Image: Tencent

Technical details

Model parameters and context
224B total (including 64B Engram parameters), 12B active per token, 1M multimodal context window
Standard API pricing (per 1M)
$0.070 input, $0.281 output and $0.003 cached input (vs $0.834 / $2.501 / $0.042 on Hy4)
Promo pricing to 30 Nov 2026
50% off on TokenHub and OpenRouter ($0.035 in / $0.1405 out / $0.0015 cache); free in WorkBuddy
Architecture and serving stack
Hybrid attention + LatentMoE + Engram (1/6 Hy4 KV cache); megakernels + 4-bit quantisation + DFly (>150 tok/s)

Release details & context

WorkBuddy turnaround gains, Yuanbao GSB uplift and public benchmark scope

Inside Tencent’s internal deployments, WorkBuddy reported that Hy4 flash shortened average office task completion time by 70% at roughly one-fifteenth the cost of Hy4 while recommending full Hy4 only for complex long-horizon software engineering, and Yuanbao recorded a 13.6% agent score gain over Hy3, a 7.56% human side-by-side (GSB) win-rate lift and a 2.52% general chat lead over GLM-5.3-Flash.

Tencent’s external evaluation suite benchmarked Hy4 flash across seven public agent tests (DeepSWE, ProgramBench, Terminal-Bench 4.0, APEX-Agents, AutomationBench, HLE and OSWorld 2.1) alongside eight internal WorkIndex tracks covering slide creation, Markdown co-writing, spreadsheet repair and video comprehension; on OpenRouter, Tencent Cloud currently serves 100% of tencent/hy4-flash traffic directly without multi-provider splitting.

Common questions

Has Tencent published open weights for Hy4 flash alongside the API launch?

Not yet; Tencent stated in its 16 November 2026 research post that it plans to release open weights for Hy4 flash in a future update.

Which workloads does Tencent still recommend routing to flagship Hy4 instead of Hy4 flash?

Tencent recommends Hy4 flash for everyday search, document, slide, frontend demo and computer-use tasks while reserving full Hy4 for long-horizon software engineering.

Tencent release timeline (2 launches tracked)

  1. Earlier4 Nov 2026
    Tencent launches multimodal Hy4 model with native computer use and 3D engine training

    Tencent released Hy4 on Tencent Cloud TokenHub at roughly $0.90 per 1M input tokens, adding multimodal perception, computer use and 3D engine workflows.

Explore the full Tencent dossier or browse all model releases.

Also in the Monday 16 November edition

Read all 15 launches on one page
  1. 01
  2. 02
  3. 03
  4. 05
  5. 06
  6. 07
  7. 08
  8. 09
  9. 10
  10. 11
  11. 12
  12. 13
  13. 14
  14. 15

Spotted an error? Email aiforsuccess.in+compass@gmail.com or tag @ai_for_success on X