Author: RadarAI Editorial
Editor: RadarAI Editorial
Last updated: 2026-11-16
Review status: Editorial review pending
Brief
速报
官方
AI动态
开源
Tencent Hunyuan releases the Hy4 flash model, achieving performance on everyday tasks comparable to Claude Haiku 5.5 with 224B total parameters, 12B activation, and 1M context, while reducing task costs by 18x, with API pricing as low as $0.070 per million input tokens [7][17][22]; meanwhile, Alibaba launches Qwen4-Max with 2.4T total parameters, and ByteDance Volcano Engine fully opens the Seedance 2.5 4K video generation API, ...
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
## 🔍 Core Insights
**Tencent Hunyuan** releases the **Hy4 flash** model, achieving performance on everyday tasks comparable to **Claude Haiku 5.5** with **224B total parameters**, **12B activation**, and **1M context**, while reducing task costs by **18x**, with API pricing as low as $0.070 per million input tokens [7][17][22]; meanwhile, **Alibaba** launches **Qwen4-Max** with **2.4T** total parameters, and **ByteDance Volcano Engine** fully opens the **Seedance 2.5 4K** video generation API. Competition among large models is shifting from parameter scale to cost-effectiveness and real-world deployment [1][21].
## 🚀 Key Updates
- **Tencent Hunyuan releases Hy4 flash** [7]: 224B total parameters, 12B activation, 1M context, blind tests comparable to Claude Haiku 5.5, task costs reduced by 18x.
- **Alibaba releases Qwen4-Max** [21]: 2.4T total parameters, native vision capabilities, supports million-token context, continuously evolving agent and coding capabilities.
- **Volcano Engine opens Seedance 2.5 4K API** [1]: 4K video generation model fully open, supports Draft mode to validate storyboards at 480P before generating the final cut.
- **ModelBest MiniCPM-OCR 1.0 gains native support** [5]: Code has been merged into vLLM and SGLang, allowing developers to deploy directly with OCR capabilities unchanged.
- **Doubao Work launches "Workbench"** [24]: Initially covers 21 professional scenarios including finance, healthcare, and education; financial investment research can link to data services such as Wind and Hithink via plugins.
- **TSMC plans to supply capacity based on customer credit limits** [4]: Customer financial data will be incorporated into demand assessment; orders exceeding limits require cash payment, and some customers must prepay to reserve capacity.
- **Musk confirms Grok API text watermarking is off by default** [2]: Watermarks are added only when legally required or proactively enabled by users.
- **Meshy open-sources ThreeRT** [10]: Brings 60 fps hardware ray tracing rendering to three.js, covering macOS, Windows, iOS, Android, and Steam Deck.
## 🔗 Sources
[1] ByteDance Volcano Engine fully opens Seedance 2.5 4K video generation model API — https://aihot.news/items/rplpo56eu6ch0j9lecwzty9je
[2] Musk confirms Grok API text watermarking is off by default — https://aihot.news/items/ik6lxh9yrue1xo80dhftwhbdx
[4] TSMC plans to supply capacity based on customer credit limits, addressing AI circular financing concerns — https://aihot.news/items/mh8sddeaod449rs7lbzkx9rob
[5] ModelBest MiniCPM-OCR 1.0 natively supports vLLM and SGLang inference frameworks — https://aihot.news/items/y68xh1e0v0t0rnhea6jn3qwsp
[7] Tencent Hunyuan releases Hy4 flash, blind tests comparable to Claude Haiku 5.5 with task costs reduced by 18x — https://aihot.news/items/zty8a2llkfo545rtr4rw22ctj
[10] Meshy open-sources ThreeRT: bringing 60 fps hardware ray tracing rendering to three.js — https://aihot.news/items/dwvtlyuew
Tencent Hunyuan releases the Hy4 flash model, achieving performance on everyday tasks comparable to Claude Haiku 5.5 with 224B total parameters, 12B activation, and 1M context, while reducing task costs by 18x, with API pricing as low as $0.070 per million input tokens [7][17][22]; meanwhile, Alibaba launches Qwen4-Max with 2.4T total parameters, and ByteDance Volcano Engine fully opens the Seedance 2.5 4K video generation API. Competition among large models is shifting from parameter scale to cost-effectiveness and real-world deployment [1][21].
🚀 Key Updates
- Tencent Hunyuan releases Hy4 flash [7]: 224B total parameters, 12B activation, 1M context, blind tests comparable to Claude Haiku 5.5, task costs reduced by 18x.
- Alibaba releases Qwen4-Max [21]: 2.4T total parameters, native vision capabilities, supports million-token context, continuously evolving agent and coding capabilities.
- Volcano Engine opens Seedance 2.5 4K API [1]: 4K video generation model fully open, supports Draft mode to validate storyboards at 480P before generating the final cut.
- ModelBest MiniCPM-OCR 1.0 gains native support [5]: Code has been merged into vLLM and SGLang, allowing developers to deploy directly with OCR capabilities unchanged.
- Doubao Work launches "Workbench" [24]: Initially covers 21 professional scenarios including finance, healthcare, and education; financial investment research can link to data services such as Wind and Hithink via plugins.
- TSMC plans to supply capacity based on customer credit limits [4]: Customer financial data will be incorporated into demand assessment; orders exceeding limits require cash payment, and some customers must prepay to reserve capacity.
- Musk confirms Grok API text watermarking is off by default [2]: Watermarks are added only when legally required or proactively enabled by users.
- Meshy open-sources ThreeRT [10]: Brings 60 fps hardware ray tracing rendering to three.js, covering macOS, Windows, iOS, Android, and Steam Deck.
🔗 Sources
[1] ByteDance Volcano Engine fully opens Seedance 2.5 4K video generation model API — https://aihot.news/items/rplpo56eu6ch0j9lecwzty9je
[2] Musk confirms Grok API text watermarking is off by default — https://aihot.news/items/ik6lxh9yrue1xo80dhftwhbdx
[4] TSMC plans to supply capacity based on customer credit limits, addressing AI circular financing concerns — https://aihot.news/items/mh8sddeaod449rs7lbzkx9rob
[5] ModelBest MiniCPM-OCR 1.0 natively supports vLLM and SGLang inference frameworks — https://aihot.news/items/y68xh1e0v0t0rnhea6jn3qwsp
[7] Tencent Hunyuan releases Hy4 flash, blind tests comparable to Claude Haiku 5.5 with task costs reduced by 18x — https://aihot.news/items/zty8a2llkfo545rtr4rw22ctj
[10] Meshy open-sources ThreeRT: bringing 60 fps hardware ray tracing rendering to three.js — https://aihot.news/items/dwvtlyuew
← Back to Updates