Tencent Hy4 Preview

Hy4 Preview: Model Introduction & Practical Guide

Hy4 preview is Tencent Hunyuan's new-generation flagship - a 770-billion-parameter MoE with 49B active parameters and a context window beyond 1M tokens - released and open-sourced on August 28, 2026. Tencent calls it a model "born for productivity," and the internal evidence backs the framing: in a blind test with 163 Tencent domain experts evaluating 203 real engineering tasks, Hy4 preview scored 2.99/4 - ahead of both GLM 5.3 (2.92) and Kimi K3 (2.94).

Here's the short version: Hy4 represents Tencent's return to the front tier of open models. It was built through co-design with Tencent's own product teams (CodeBuddy for coding, WorkBuddy for office work), trained with expert data from software engineering, gaming, finance, and security, and evaluated on real tasks rather than curated benchmarks. The model is particularly strong in long-horizon software development, office analysis and financial workflows, one-prompt playable game prototypes, and scientific research assistance. It is open-sourced, priced at ¥6/¥18 per million input/output tokens (cache hits ¥0.3), and available across Tencent's products and via Tencent Cloud and the model API. The caveat is the label: this is a preview, with the next batch of Hy4 models promised "soon."

This guide covers model overview, core features, technical specifications, capability comparison, core advantages, recommended use cases, example prompts, and selection recommendations.

Quick Facts

Table of Contents

1. Model Overview
2. Core Features
3. Technical Specifications
4. Capability Comparison
5. Core Advantages
6. Recommended Use Cases
7. Example Prompts
8. Selection Recommendations
9. FAQ
10. Sources & Further Reading

1. Model Overview

Tencent's Hunyuan line has iterated quickly through 2026: infrastructure rebuilt from late January, Hy3 preview in April, Hy3 formal release in July, and now Hy4 preview in late August - each cycle explicitly framed as "preview first, formal version follows," with real-world feedback feeding the next iteration.

Hy4 preview's scale-up is substantial: 770B total parameters with 49B active, context beyond 1M tokens, and expanded pretraining and post-training data. But the more interesting story is how it was built. Tencent describes a co-design process with its own productivity products - CodeBuddy and WorkBuddy - and with high-quality data contributed by top experts across software engineering, gaming, finance, and security. That methodology shows in the capability profile: the improvements are concentrated in work that produces deliverables (code, decks, analyses, reports) rather than in abstract reasoning benchmarks.

The self-improvement detail is worth pausing on. Tencent states that Hy4 preview participated in optimizing its own training pipeline for the first time - proposing methods, running experiments, iterating on results, with generated code and logs feeding the next round. It also autonomously analyzed inference bottlenecks and optimized operator fusion and communication, raising end-to-end throughput by 31.8% versus baseline. This is a preview-stage demonstration of recursive self-improvement in an industrial training pipeline, and Tencent flags it as an early loop rather than a finished capability.

Distribution is broad by design: Hy4 preview shipped simultaneously in Tencent's consumer and productivity products (Yuanbao, ima, CodeBuddy, WorkBuddy) and is available for API integration through Tencent Cloud's Tokenhub and the model API - which makes it one of the few Chinese flagship models reachable through a Western aggregator on day one.

2. Core Features

770B/49B MoE at 1M+ context. Flagship-scale capacity with efficient active-parameter routing, and a context window that swallows entire codebases, document archives, and long conversations.

Long-horizon software engineering. Improved understanding, planning, debugging, and validation for extended development tasks, with particular gains in frontend visual quality and interaction polish.

Office analysis and finance. Stronger comprehension of complex office environments, financial analysis, and cross-file collaboration - covering the full path from information processing to delivered documents, spreadsheets, and presentations.

One-prompt game prototypes. Enhanced ability to generate playable prototypes directly from a single requirement, with competent use of game engines and iterative refinement through multi-turn interaction.

Scientific research support. Measurable improvements in AI R&D, molecular dynamics simulation, condensed matter physics, and foundational mathematics scenarios.

Recursive self-improvement (early). Participated in optimizing its own training methods, data strategy, evaluation systems, and low-level operators - plus autonomous inference-infrastructure optimization (+31.8% throughput).

Open weights and broad access. Open-sourced on release; available in Tencent products, the API (Tokenhub, the model API), and self-hosted deployments.

3. Technical Specifications

Cost note. At ¥6/¥18 per million tokens with cache hits at ¥0.3, long-context agent workloads with stable prefixes (codebases, document sets) become materially cheaper than uncached equivalents - the pricing structure rewards exactly the workflows Hy4 targets.

4. Capability Comparison

Honest read on the comparison. The blind-test figures come from Tencent's internal evaluation with its own experts on its own task set - credible as a directional signal of productivity-task quality (blind scoring removes brand bias), but not a substitute for independent benchmarking. What the comparison does establish is intent: Hy4 preview was tuned to compete with GLM 5.3 and Kimi K3 on exactly the productivity tasks that Chinese enterprises are buying AI for in 2026, and Tencent believes it now leads that specific contest.

5. Core Advantages

1. Productivity-task specialization. Training data co-built with experts in software, gaming, finance, and security targets the work enterprises actually automate.
2. Open weights at release. No waiting period - deploy or integrate immediately, with an open-source hedge against vendor risk.
3. Genuine long-context economics. 1M+ context with cache hits at ¥0.3 per million tokens makes sustained long-context agents affordable.
4. Broad access paths. Consumer products, Tencent Cloud, and the model API cover everything from casual use to Western-market integration.
5. Self-improvement evidence. A 31.8% throughput gain achieved through autonomous analysis is an unusual, concrete datapoint about the model's agentic quality.
6. Rapid iteration cadence. Preview-to-formal cycles measured in months, with the next Hy4 models already announced.

6. Recommended Use Cases

• Long-horizon software development: multi-day engineering tasks, frontend quality work, debugging and validation pipelines through CodeBuddy or API.
• Office and finance workflows: analysis, cross-file collaboration, and deliverable generation (documents, spreadsheets, presentations).
• Game prototyping: one-sentence-to-playable-prototype generation and iterative refinement in game engines.
• Scientific computing assistance: research workflows in AI R&D, molecular dynamics, condensed matter physics, and mathematics.
• Enterprise knowledge work at scale: long-context document processing with cache-friendly repeated queries.
• Self-hosted deployments: organizations deploying open-weight flagships on their own infrastructure.

7. Example Prompts

1. Long-horizon engineering task

2. Financial analysis deliverable

3. One-prompt game prototype

4. Scientific research assistance

5. Long-context document workflow (cache-friendly)

8. Selection Recommendations

Choose Hy4 preview if:

• Your workload is productivity-oriented: code, office analysis, financial workflows, game prototypes, research support.
• You want open weights from a flagship-class model with immediate availability.
• You need 1M+ context with cache pricing that makes long-context agents sustainable.
• You want integration through Western aggregators (the model API) without building China-market infrastructure.
• You are already in Tencent's ecosystem (CodeBuddy, WorkBuddy, Yuanbao, ima).

Consider alternatives if:

• You need the largest open model by parameter count (Kimi K3's 2.8T) or its specific visual-engineering workflows.
• You require independent (non-vendor) benchmark validation before committing - preview-stage models warrant your own evaluation regardless.
• You need guaranteed long-term API stability today (preview status implies churn; the formal release will follow).

Sources & Further Reading

• Tencent releases and open-sources Hy4 preview - Tencent (official)
• Hy4 preview - Tencent Hunyuan research page
• Hy4-preview - GitHub
• Hy4 preview analysis - Zhihu (Chinese)
• Hy4 - Baidu Baike (Chinese)

Hy4 preview is a preview-stage model: pricing, availability, and capabilities may change with the formal release and subsequent Hy4 models. Internal blind-test figures are vendor-published - run your own task suite before making adoption decisions.