Tencent Hy4 Preview API

Tencent Hy4 Preview is a next-generation large language model released and open-sourced by Tencent's Hunyuan team on August 28, 2026, positioned as "built for productivity." It uses a MoE architecture with 770B total parameters and 49B activated parameters, and a context window exceeding 1M tokens. It demonstrates outstanding capabilities across coding, office work, and scientific research. In Tencent's internal blind evaluation with 163 experts and 203 engineering tasks, it scored 2.99/4, slightly ahead of GLM 5.3 (2.92/4.00) and Kimi K3 (2.94/4.00). Model weights are open-sourced with a permissive commercial license, launching simultaneously across WorkBuddy/CodeBuddy, Yuanbao, ima, and other Tencent products.

On iCreat, developers call hy4-preview through an OpenAI Chat Completions-compatible API. Pricing: input $0.834/million tokens, output $2.501/million tokens, cached read $0.042/million tokens.

Model Positioning

Hy4 Preview is the first public version of the Hunyuan Hy4 series, expanding significantly in model size, context length, and data volume over Hy3. It is positioned as a production-grade productivity model, focused on software engineering, office analytics, game development, and scientific research. The model participated for the first time in the automated optimization of its own training methods, data strategies, evaluation frameworks, and low-level operators, establishing an early-stage recursive self-improvement loop. It also autonomously optimized its inference system bottlenecks, increasing end-to-end throughput by 31.8% over the baseline. WorkBuddy/CodeBuddy offer a two-week limited free trial.

Core Capabilities

Software engineering

Enhanced understanding, planning, debugging, and validation for long-context development tasks, improving front-end visual aesthetics and interaction quality. Ranked 5th in Arena coding tests, 3rd among open-source models.

Office analytics

Significantly improved complex office environment understanding and financial analysis capabilities, with optimized data analysis and cross-document collaboration — supporting the full workflow from information processing through documents, spreadsheets, and presentations.

Game development

Generates a playable prototype from a single natural-language request, proficient with game engines. Developers can continue refining complex game projects through multi-turn interactions.

Scientific research

Improved understanding, reasoning, and problem-solving for complex research problems, with notable advances in AI R&D, molecular dynamics simulation, condensed-matter physics, and fundamental mathematics.

Recursive self-improvement

Participated for the first time in automated optimization of training methods, data strategies, evaluation frameworks, and low-level operators — proposing approaches, running experiments, and iterating based on results, forming an early-stage recursive self-improvement loop. Also autonomously analyzed inference system bottlenecks, optimizing operator fusion and communication, increasing end-to-end throughput by 31.8%.

Cached-read discount

Cached reads billed at $0.042/million tokens (approximately one-twentieth of the standard $0.834 input price), significantly reducing costs for repeated-context workloads.

Pricing

Total cost = input tokens × $0.834/million + output tokens × $2.501/million (cached reads at $0.042/million).

Note: Tencent Cloud TokenHub domestic pricing is ¥6/million input, ¥18/million output, cache hit ¥0.3/million (CNY); the USD prices above are iCreat platform prices, matching Tencent's official overseas pricing.

Use Cases

• Software engineering: long-context development task understanding, planning, debugging, and validation; front-end visual quality
• Office and financial analysis: cross-document collaboration, data analysis, full workflow from information processing to documents/spreadsheets/presentations
• Game development: playable prototype from a single prompt, multi-turn refinement of complex projects
• Scientific research: AI R&D, molecular dynamics simulation, condensed-matter physics, fundamental mathematics
• Cost-sensitive high-frequency calls: cached reads at only $0.042/million tokens — ideal for RAG and multi-turn conversations with shared system prompts

Model Comparison

Hy4 Preview vs. Hy3

Hy4 Preview vs. Hunyuan TurboS and Hunyuan T1

Note: Hunyuan TurboS and T1 were taken offline on June 22, 2026; Hy4 Preview is Hunyuan's latest model.

Why Choose Hy4 Preview?

• 770B open-source flagship: top-tier among open-source models, Arena coding 5th (3rd open-source)
• 1M+ context: supports entire code repositories and massive long documents in a single input
• Recursive self-improvement: first Hunyuan model to participate in its own training optimization; inference throughput autonomously improved 31.8%
• Full productivity coverage: software engineering, office analytics, game development, and research — all enhanced
• High value: cached reads at only $0.042/million; output at $2.501 below same-tier flagships
• Two-week free trial: WorkBuddy/CodeBuddy limited free access; Hy3 free period extended to September 30

Specifications

Architecture

Hy4 Preview uses a MoE architecture with 770B total parameters and 49B activated per token, with a context window exceeding 1M tokens. The model expanded significantly in size, context length, and data volume over Hy3, with advances in both pre-training and post-training leading to a major leap in overall intelligence.

The model participated for the first time in the automated optimization of its own training methods, data strategies, evaluation frameworks, and low-level operators — proposing approaches, running experiments, and iterating based on results, forming an early-stage recursive self-improvement loop. It also autonomously analyzed inference system bottlenecks, optimizing operator fusion and communication, increasing end-to-end throughput by 31.8% with consistent gains across different context lengths and concurrency levels.

Since the infrastructure rebuild in February 2026, Hunyuan has iterated a major version approximately every two months (Hy3 Preview April → Hy3 July → Hy4 Preview August). The next batch of Hy4 series models is expected soon.

Notes

• The context window exceeds 1M tokens and is a shared request space; reserve room for system instructions, conversation state, and the final answer
• When thinking mode is enabled, responses include reasoning token consumption; for latency- and cost-sensitive high-throughput scenarios, use reasoning_effort: "low"
• Cached-read discount applies only when the input prefix hits the model cache
• This is a Preview version; a formal release is expected to follow — monitor Hunyuan official channels
• Agent loops that repeatedly resend large context blocks should leverage caching: cached reads are only $0.042/million tokens

FAQ

What is the difference between Hy4 Preview and Hy3?

Hy4 Preview is the next-generation model (released August 28, 2026, 770B total / 49B activated, 1M+ context), comprehensively surpassing Hy3 (295B / 21B, 256K context) in productivity tasks. Hy3 is the official release from July 6, positioned for high value; Hy4 Preview is a productivity flagship, Arena coding 5th (3rd open-source), first to support recursive self-improvement, with throughput autonomously improved 31.8%. Both are open-source.

Is Hy4 Preview the official release?

No. Hy4 Preview is the first public version (preview) of the Hy4 series, designed to collect real-world user feedback. Following Hunyuan's roughly two-month iteration cadence, the formal Hy4 release is expected soon.

What is Hy4 Preview's context window?

It exceeds 1M tokens, supporting entire code repositories and massive long documents in a single input.

How do I enable thinking mode?

Set "thinking": {"type": "enabled"} in the request body, optionally adjusting reasoning_effort (low/medium/high). Lower or disable thinking on simple tasks to save latency and cost; raise the tier for complex tasks.

How is cached-read billing calculated?

Cached input is billed at $0.042/million tokens instead of the standard $0.834/million input price — a roughly 20× reduction. Ideal for RAG and multi-turn conversations with shared system prompts.

Did the model participate in its own training?

Yes, for the first time in Hunyuan's history. Hy4 Preview participated in the automated optimization of training methods, data strategies, evaluation frameworks, and low-level operators — proposing approaches, running experiments, and iterating based on results, forming an early-stage recursive self-improvement loop. It also autonomously optimized its inference system, increasing end-to-end throughput by 31.8%.