Skip to main content

Hy4 Preview

Overview

Hy4 Preview is an open-weight Mixture-of-Experts language model released by the Tencent Hy Team on August 28, 2026. It is an early Hy4-series release with 770 billion backbone parameters, 49 billion activated parameters per token, and a 1M-token context window. On B.AI, use the model ID hy4-preview.

Key Features

  • 770B Sparse MoE Architecture: Uses 78 backbone layers, 256 routed experts and one shared expert per MoE layer, and activates the top eight routed experts plus the shared expert for each token. A separate native MTP layer adds 10B total and 0.7B activated parameters for speculative decoding.
  • Long-Context API Limits: Supports a 1M-token context window, up to 960K input tokens, and up to 64K output tokens.
  • Preserved Deep Reasoning: Thinking is enabled by default. The API exposes reasoning_content, which should be returned with the complete assistant message during multi-turn tool workflows.
  • Agent API Support: Supports structured output, function calling, cached input, and OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages interfaces.
  • Productivity-Oriented Validation: In Tencent's blind internal evaluation, 163 experts rated outputs on 203 engineering tasks. Hy4 Preview averaged 2.99/4.00, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3.

Best Use Cases

  • Software Engineering Agents: Long-horizon planning, implementation, debugging, verification, frontend work, and multi-step tool use across large repositories.
  • Office and Data Workflows: Cross-document analysis, financial modeling, data analysis, and production of documents, spreadsheets, and presentations.
  • Game Development: Playable-prototype generation, game-engine workflows, and iterative refinement over multiple turns.
  • Scientific Research: Tool-assisted analysis and problem solving in areas such as AI research, molecular dynamics, condensed-matter physics, and mathematics.
  • Private or Customized Deployment: Open-weight deployments that require local serving, quantization, or task-specific fine-tuning and can accommodate the model's substantial compute requirements.

Capabilities and Limitations

CapabilityDescription
ReasoningSupports deep reasoning with preserved thinking. The documented reasoning-effort levels are low and high, with high as the default for hy4-preview tool workflows.
Creative WritingSupports general text generation and Chinese content creation.
MultimodalText input and text output.
Response SpeedTencent has not published a guaranteed generation rate. The official release reports a 31.8% throughput gain from model-assisted inference-system optimization relative to Tencent's internal baseline.
Context Window1M tokens.
Maximum Input960,000 tokens.
Max Output64,000 tokens.
Tool UseFunction calling, structured output, automatic tool selection, cached input, and preserved reasoning across tool calls.
MultilingualTencent documents Chinese and English use.

Known Limitations

  • Tencent identifies this release as an early version of Hy4 rather than a final Hy4 model.
  • The official model card reports that Hy4 Preview can spend longer than necessary reasoning through complex tasks and can over-verify its own work.
  • Multi-turn tool integrations should preserve and replay the full assistant message, including reasoning_content; omitting it can degrade the documented continuation workflow.
  • Hy4 Preview does not support native image, audio, or video understanding.
  • Tencent has not published a model-specific knowledge cutoff, guaranteed API throughput, or complete language-support list.

Pricing

ModelInput (Credits/Token)Cache Write (Credits/Token)Cache Read (Credits/Token)Output (Credits/Token)Web Search (Credits/Use)
Hy4 Preview0.8340.8340.0422.501-
Pricing note

Prices shown in the documentation are B.AI standard reference prices for base billing purposes. B.AI may provide lower actual usage costs through top-up bonuses and account benefits. Specific prices, bonus Credits, and account benefits are subject to the platform display and final billing records.