The whole dayFri 20 Nov 2026
AI Compass
WeeklyArchiveAbout

Tencent open-sources 770B Hy4 multimodal MoE and Hy4-FP8 weights

Live on Hugging Face, ModelScope and GitCode on 20 November 2026 alongside pre-trained Hy4-base, a 3.3B DFly2 drafter and 1M context under the Tencent Hy License.

Luminous blue-and-cyan Tencent Hy swirling orb suspended above a crystal sparse-attention cube routing glowing optical fibres into dark GPU compute modules
Image: AI Compass illustration (Tencent marks)

Tencent’s Hy Team published open weights for its flagship Hy4 instruct model (tencent/Hy4), FP8-quantised Hy4-FP8, pre-trained Hy4-base and an end-to-end test GGUF build on 20 November 2026 across Hugging Face, ModelScope and GitCode. The release fulfils the open-weight pledge from Tencent’s 4 November Hy4 API launch and follows 220B open-weight Hy4-flash. Its 78-layer language backbone pairs one dense layer with 77 Mixture-of-Experts layers totalling 770B parameters, activating top-8 of 256 routed experts plus one shared expert for 49B active text parameters and 1.23B vision parameters per token across a 1,048,576-token context window.

Gated DeepSeek Sparse Attention, DFly2 zero-KV drafting and serving images

Hy4 combines Multi-head Latent Attention with Gated DeepSeek Sparse Attention, selecting the top 2,048 tokens per step and reusing sparse indices across layers via IndexCache alongside four-stream identity Hyper-Connections. Built directly onto the backbone is DFly2, a five-layer, 3.3B-parameter speculative-decoding drafter trained with AngelSpec that shares Hy4’s KV cache and top-2,048 sparse positions so each draft token attends to roughly 2,000 keys with zero extra KV memory.

DFly2 speculative decoding architecture diagram showing target DSA layers sharing top-k sparse KV caches and indices with three drafter transformer blocks and gated hidden fusion
DFly2 reuses Hy4’s top-2,048 sparse KV cache and indexer positions across draft layers at constant compute and zero extra KV memory.Image: Tencent

Tencent ships prebuilt multi-arch Docker containers (vllm/vllm-openai:hy4 and lmsysorg/sglang:hy4), its AngelSlim compression suite and the Hy4 System Card under the Tencent Hy License, which requires separate commercial licensing above $50 million in annual MaaS revenue or 100 million monthly active users.

Twelve-panel bar chart comparing Hy4 against Hy4 preview, GLM 5.3, DeepSeek-V4.1-Flash-1017, Kimi K3.1 Pro, Claude Opus 5.5 and GPT-6.1 Sol across coding, agent and OSWorld 2.1 benchmarks
Hy4 scores 78.9 on OSWorld 2.1, 70.0 on E-Bench 2.0 and 44.4 on Terminal Bench 4.0 against open-weight and proprietary peers.Image: Tencent

Technical details

Checkpoints published
tencent/Hy4 (instruct), tencent/Hy4-FP8, tencent/Hy4-base and end-to-end test GGUF
Backbone and active parameters
770B language backbone (78 layers: 1 dense + 77 MoE); 49B text + 1.23B vision active; 3.3B DFly2 drafter
Attention and serving containers
MLA + Gated DSA (top-2,048 IndexCache); vllm/vllm-openai:hy4 and multi-arch lmsysorg/sglang:hy4
Licence and hosted API rates
Tencent Hy License ($50M MaaS or 100M MAU threshold); API at $0.834 input, $0.042 cached, $2.501 output per 1M

Release details & context

Reasoning modes, prefill-decode disaggregation and safety red-teaming

When served on vLLM or SGLang, Hy4 defaults reasoning_effort to high for deep chain-of-thought math and coding, while passing chat_template_kwargs reasoning_effort set to no_think returns direct answers at recommended sampling settings of temperature 0.9 and top_p 1.0. Because DFly2 skips prefill entirely, prefill-to-decode disaggregation transfers only one final-layer hidden state between nodes.

Across domain benchmarks, Hy4 ranked first among open-weight models on TenureTrail (123 scientific tasks across seven disciplines) and cut internal financial hallucination rates from 15.76% on Hy4 preview to 5.45%, while scoring 30.4 out of 100 on Hy-BioExperimentBench and passing offensive cryptanalysis and binary-obfuscation audits by Tencent Xuanwu Lab and Keen Security Lab.

Common questions

How does Hy4’s DFly2 speculative drafter avoid extra KV-cache memory on 1M-token contexts?

Each of DFly2’s five draft layers shares the backbone’s KV cache and Gated DSA top-2,048 selected token indices directly, capping draft attention at roughly 2,000 keys without allocating a separate KV store.

Who needs a separate commercial agreement under the Tencent Hy License?

Organisations operating Model-as-a-Service platforms above $50 million in annual revenue or products exceeding 100 million monthly active users must request a separate commercial licence from Tencent.

Weekly radarAI Compass Weekly · 16 to 20 Nov

Editor’s weekly perspective

Open-weight MoEs, robotics and RL harnesses: Publishes 770B flagship Hy4, Hy4-FP8 and Hy4-base checkpoints (49B active text) with a 3.3B DFly2 drafter.

Weekly benchmark note: 770B / 49B active multimodal MoE + 3.3B DFly2 drafter, 1M context, open Hy License + FP8

Read the 16 to 20 Nov Weekly Edition

Sources

  1. Primarytencent/Hy4 on Hugging Facehuggingface.co/tencent
  2. Officialtencent/Hy4-FP8 on Hugging Facehuggingface.co/tencent
  3. Officialtencent/Hy4-base on Hugging Facehuggingface.co/tencent
  4. OfficialTencent Hy Research: Introducing Hy4hy.tencent.ai

Tencent release timeline (4 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.

  2. Earlier16 Nov 2026
    Tencent launches 224B Hy4 flash multimodal MoE at $0.070 input

    Tencent launches its 224B-total, 12B-active Hy4 flash multimodal model at $0.070 input and $0.281 output, at one-twentieth of Hy4’s task cost.

  3. Earlier18 Nov 2026
    Tencent releases open-weight 220B Hy4-flash models with 1M context

    Tencent open-sources its 220B-total, 12B-active multimodal Hy4-flash, Hy4-flash-FP8 and Hy4-flash-base models on Hugging Face.

Explore the full Tencent dossier or browse all open source releases.

Also in the Friday 20 November edition

Read all 13 launches on one page
  1. 01
  2. 03
  3. 04
  4. 05
  5. 06
  6. 07
  7. 08
  8. 09
  9. 10
  10. 11
  11. 12
  12. 13

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

Email Edition · Early Access

Get the 5-minute AI briefing

Daily and Sunday Weekly inbox editions open soon. Join the early-access list.

Zero spam · Privacy

! Please enter a valid email address (for example, you@domain.com).

Morning Edition Alerts

Get a 1-click alert at 11:00 IST

One quiet browser alert when each daily or Sunday Weekly edition goes live between 11:00 and 12:00 IST. No email required. Privacy