Tencent has released Hy4 Preview, an open-weight model under the Apache 2.0 license, featuring a Mixture-of-Experts architecture with 770B total parameters, 49B active parameters, and a 1M context window. The model is designed for productivity and is available for testing, with resources on Hugging Face and GitHub.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Tencent releases Hy4 Preview as an open-source model with 770B parameters and 1M context, focused on real-world work output. Benchmarks show domestic models in a tight race, but ARC-AGI-2 trails Gemini 3.1 Pro by 24.5 points, making abstract reasoning the biggest shortfall.
Tencent open-sourced Hy4 Preview on August 28, 2026, with 770B parameters, 49B active parameters, and 1M context length, a significant upgrade from Hy3's 295B/21B and 256K. The company calls it 'open-source frontier' and emphasizes 'built for productivity.'
Notably, this is only a preview version; following Hy3's iteration pace, the full release is expected to be stronger.
The report provides two sets of benchmarks: 10 against global frontier models (GPT 5.5, Gemini 3.1 Pro, Claude Opus 4.8, and three domestic models) and 18 against domestic models (DeepSeek V4, GLM 5.3, Kimi K3). The conclusion is that domestic models are in a tight lead (13/18 firsts), while globally they achieve single breakthroughs but overall catch-up (2/10 firsts).
Against domestic models, Hy4 takes the top spot, but in the 5 categories it didn't win, 4 were lost to GLM 5.3, with margins mostly within 1 point. In a blind test with 163 internal experts, Hy4 scored 2.99 vs GLM 5.3's 2.92, a net win rate of only about 6 percentage points. Kimi K3 lags half a step behind overall, while DeepSeek V4 falls behind notably in agentic tasks (Terminal-Bench 62.2, BrowseComp 64.0).
Against global frontier models, strengths are concentrated in agentic and search: Terminal-Bench 3.0 scores 77.3 (about 1.7 points above GPT 5.5) and BrowseComp scores 83.6 (about 2.2 points above GPT 5.5), both world firsts. The largest generational improvements are also in these two plus ARC-AGI-2, with gains of +18.9, +16.9, and +16.4 percentage points respectively.
On ARC-AGI-2, Hy4 scores only 52.6, trailing Gemini 3.1 Pro (77.1) by 24.5 points and GPT 5.5 (70.7) by 18 points, the most glaring gap in all data. Domestic models overall (GLM 48.3, Kimi 45.4, DeepSeek 35.0) still have a generational distance from US frontier models, indicating a clear deficiency in reasoning that requires generalization rather than pattern matching.
Math and science reasoning (AIME, IMOAnswerBench, GPQA) enter the first tier, with gaps of 1–4 points, a catchable range.
Hy4 Preview's training data construction is unique: it collaborates with Tencent's internal software engineers, game developers, financial analysts, and security experts, building training data around their actual delivered work. This approach converts internal industry scenarios into training signals, a differentiated resource for Tencent compared to pure model companies.
The four capability domains include: software engineering (including front-end 'visual taste and interaction quality,' showcased with a Three.js museum case), office analysis (multi-file messy context to deliverable documents/tables/PPT, with an expense review case emphasizing multi-version policy applicability judgment and fraud detection), game development (single prompt to generate playable prototypes, supporting Unity/Unreal multi-round iteration), and scientific research (AI research, molecular dynamics, condensed matter physics, pure mathematics).
The report also reveals that Hy4 Preview managed multiple Codex sessions in parallel, playing a 'researcher' role in a small-model post-training task, ultimately surpassing Codex's solo exploration on all 8 benchmarks, which is seen as a noteworthy signal.
This report is based on a repost by X user meng shao, citing Tencent Hy's official account release and benchmark report. All data comes from that source and has not been independently verified. Some conclusions (e.g., 'full version will be stronger') are inferences.
Hy4 Preview leads domestic models but in a tight race, with strengths in agentic and search, and a clear weakness in abstract reasoning. Its training data strategy is worth attention, but the full version's performance and real-world application effects remain to be seen.
Primary report
Primary source
Also covered by 1 sources