Tencent Hy4 Preview Model Release

Tencent released the Hy4 Preview open-source model on August 28, 2026. The model features 770B total parameters with 49B active parameters and a 1M-token context window, compared to Hy3's 295B/A21B and 256K. Training data was co-constructed with Tencent's internal software engineers, game developers, financial analysts, and security experts, turning real industrial deliverables into training signals. Four capability domains are highlighted: software engineering (including visual taste and interaction quality), office analysis (multi-file messy context to deliverable documents/sheets/PPT, with fraud detection), game development (single-prompt playable prototypes, Unity multi-round iteration), and scientific research (AI research, molecular dynamics, condensed matter physics, pure math). Hy4 Preview once managed multiple Codex sessions in parallel, acting as a researcher to optimize a post-training task and surpassing Codex alone on all 8 benchmarks. Evaluation against 10 global frontier models shows 2 first places (Terminal-Bench 3.0: 77.3, BrowseComp: 83.6), while against 18 domestic models it leads in 13, with GLM 5.3 close behind. The largest gap to global models is abstract reasoning (ARC-AGI-2: 52.6 vs Gemini 3.1 Pro's 77.1). On Arena WebDEV, Hy4 Preview ranked 5th, similar to GLM-5.3Flash. A separate commentary emphasized that expert high-quality data co-construction is critical for model performance, warning against algorithm arrogance.
Key insights
Hy4 Preview brings a massive parameter jump and 1M context
Tencent's Hy4 Preview scales to 770B total parameters with 49B active, quadrupling Hy3's 295B/A21B, and extends context length from 256K to 1M tokens. This positions it among the largest open-source models and directly targets long-document and long-horizon agent tasks.
Training data built from internal industrial workflows is Tencent's differentiator
Unlike models trained solely on public data, Hy4 Preview's training set was co-constructed with Tencent's own software engineers, game developers, financial analysts, and security experts, using their actual work deliverables. This turns internal industrial scenarios into training signals, a resource advantage over pure model companies.
Agentic and search capabilities are the standout strengths
Hy4 Preview achieved global firsts on Terminal-Bench 3.0 (77.3, +1.7 over GPT 5.5) and BrowseComp (83.6, +2.2 over GPT 5.5). These two benchmarks plus ARC-AGI-2 saw the largest generational improvements (+18.9, +16.9, +16.4 percentage points), indicating post-training resources were heavily allocated to long-cycle agent tasks.
Domestic leadership is narrow, not a generational gap
Hy4 Preview leads in 13 of 18 domestic benchmarks, but lost 4 to GLM 5.3 (IMOAnswerBench, GPQA-Diamond, SWE-bench Verified, tau3-bench) with margins mostly under 1 point. A 163-expert blind test showed a net win rate of only about 6 percentage points over GLM 5.3 (2.99 vs 2.92). Kimi K3 trails by a larger margin, and DeepSeek V4 lags significantly in agentic tasks.
Abstract reasoning remains the biggest gap to global frontier models
On ARC-AGI-2, Hy4 Preview scored only 52.6, trailing Gemini 3.1 Pro (77.1) by 24.5 points and GPT 5.5 (70.7) by 18 points. This stark gap shows that domestic models still have a generational distance in generalization-based reasoning, even as math and science reasoning enter the first tier with gaps of 1–4 points.
Expert data quality is the key driver of model progress
A commentary on the release stressed that the sentence 'through high-quality data co-construction with top experts... Hy4 preview has made significant progress' is the most important takeaway. It draws a parallel to DeepSeek's practice of researchers spending most of their time labeling data, arguing that algorithm arrogance without data focus slows progress.
Hy4 Preview ranks 5th on Arena WebDEV, comparable to GLM-5.3Flash
On the Arena WebDEV leaderboard, Hy4 Preview placed 5th, with performance similar to the recently released GLM-5.3Flash. This provides an additional external validation point alongside the official benchmark results.
Hy4 Preview acted as an autonomous researcher coordinating multiple Codex sessions
A disclosed experiment showed Hy4 Preview managing multiple Codex sessions in parallel for a small-model post-training task, judging directions, organizing experiments, and coordinating optimization across 8 evaluation targets, ultimately outperforming Codex alone on all benchmarks. This demonstrates early autonomous research coordination capabilities.