HOT RADAR
llmHEAT 0.0°

AI CURATED HOT STORY · 11/4/2026

Tencent Unveils 770B Hy4 with Vision and Computer Use

11 reports archived7 independent sourcesupdated 11/9/2026, 16:44:00
Synthesis & Latest Updates
7 Sources Cross-Validated

Tencent has officially launched Hunyuan Hy4, a new large language model with 770B total parameters (49B activated), up from under 300B in Hy3. It introduces native visual understanding and computer use capabilities for the first time, and the company claims it ranks among the strongest open-source models. Weights are expected to be released soon with support from the vLLM/SGLang ecosystems.

Across sources, Hy4 shows significant improvements in instruction following, long-context understanding, and long-horizon agent tasks, supporting a 1M token context and image-text input. It scores 35 on the Artificial Analysis Intelligence Index (second among Chinese models) and ranks 14th on Agent Arena with a cost of just $0.55 per task, offering strong value. Its vision can locate ~2-meter components in 2.8-billion-pixel aerial images, and computer use covers CLI and GUI, with specialized training for Blender and CAD. However, Tencent admits it still lags top closed-source models on hard math and long-horizon coding, and its AA-Omniscience hallucination rate is 89%.

Hy4 is already integrated into Tencent products like WorkBuddy, CodeBuddy, Yuanbao, and WeCom, as well as OpenRouter. First-day demand surged, prompting a tenfold expansion of WorkBuddy's inference cluster. Pricing remains at ¥6 per million input tokens and ¥18 per million output tokens, with a cheaper Hy4 Flash planned. This marks Tencent's entry into the top tier of open-source LLMs, but the actual ecosystem impact will depend on the forthcoming weight release.

LATEST/First-day demand for Hy4 surged, prompting a tenfold expansion of WorkBuddy's inference cluster and increased free quotas. Model weights and system card will be released soon, with Tencent Cloud offering a 50% discount. Tencent confirms Hy4 has been used to improve training of other models.

Coverage timeline

Total 11 reports · Latest first
  1. 量子位官网T2·72 pts
    • Hy4 expands pretraining and post-training scale versus the preview and adds vision understanding for the first time
    • Second among Chinese models on Artificial Analysis' index and first among open models on Agent Arena, at about $0.78 per task
    • First-day traffic spikes triggered a tenfold inference scale-up, signaling real demand and capacity pressure
  2. Hacker News AIT2·72 pts
    • Hy4 adds first-time visual understanding and computer-use abilities spanning CLI and GUI interaction, chart analysis and math visual reasoning
    • Tencent claims gains in instruction following, long context and long-horizon agentic tasks, but provides no concrete benchmark figures
    • Weights and system cards are due within days, bundled with limited-time free access and 50% off cloud pricing
  3. Artificial Analysis (@ArtificialAnlys)T1·72 pts
    • Hy4 scores 35 on the Intelligence Index, tied with Grok 4.7 and #2 among China-based labs
    • Leads Chinese models on FrontierCode 2.0 at 31%, close to Claude Fable 5.1's 32%
    • Hallucination rate of 89% on AA-Omniscience, far above Kimi K3.1 Pro's 37%
    • 1M-token context, text and image input, priced at $0.834/$2.501 per 1M tokens
  4. InfoQ 中文T2·78 pts
    • Parameters jump from under 300B in Hy3 to 770B, with Tencent claiming a first-tier open-model position and weights due soon
    • First-time vision understanding and Computer Use enable cross-application, long-horizon tasks via CLI and GUI
    • Pricing unchanged from Preview at RMB 6/18 per million input/output tokens with RMB 0.3 cache hits, plus a cheaper Hy4 Flash in the pipeline
  5. IT之家 · 智能时代T2·78 pts
    • Adds vision for fine-grained image processing, complex chart analysis and math visual reasoning
    • Supports computer use for the first time, combining CLI and GUI, with targeted training for Blender and CAD
    • Weights are slated for open release with vLLM/SGLang ecosystem support and a forthcoming system card
    • Tencent concedes a notable gap to top closed models on hard math, long-horizon coding and visual QA benchmarks