Hy4 preview

Hy4 preview is a new-generation Mixture-of-Experts flagship from Tencent's Hy Team, released on 2026-08-27. It uses 770B total parameters of which 49B are activated per token, a 1M-token context window, and Apache 2.0 licensing, putting it at the open-source frontier. Tencent ran a blind side-by-side evaluation with 163 internal experts rating 203 engineering tasks, and Hy4 preview edged out both GLM 5.3 (2.99 vs 2.92) and Kimi K3 (2.99 vs 2.94).

Key Features

• 7:1 activation ratio: 770B total, 49B activated per token.
• Gated DeepSeek Sparse Attention (DSA): Extends the DeepSeek/GLM sparse-attention line, paired with IndexCache for cross-layer sparse index reuse.
• iHC residual pathway: Identity Hyper-Connections expand inter-layer information flow.
• 1M-token context: Handles very long-horizon engineering and analysis tasks in one pass.
• Native MTP layer: A built-in multi-token prediction layer (10B params, 0.7B activated) speeds up speculative decoding.
• Productivity-focused training data: Built around the work of Tencent software engineers, game developers, financial analysts, and security experts.

Use Cases

Who Should Use This Tool?

• Long-horizon software engineering: Understanding, planning, debugging, and verifying large multi-file development tasks.
• Office and analysis: Turning messy multi-file context into documents, spreadsheets, and presentations with better precision.
• Game development: Turning a single prompt into a playable prototype and iterating with game engines.
• Scientific research: Stronger reasoning on hard questions across AI research, molecular dynamics, condensed-matter physics, and pure math.

Problems It Solves

1. Context overflow on big tasks: A 1M token window covers an entire project or report in one pass.
2. Sparse attention cost: Gated DSA plus IndexCache reduces the compute of long-context inference.
3. Task drift on long horizons: The activation budget and reasoning focus keep a model on track across multi-turn engineering work.

Pricing

Advantages & Unique Selling Points

• Open frontier scale: 770B/49B at Apache 2.0 with a 1M context.
• Beats peers on blind engineering eval: 2.99 vs GLM 5.3's 2.92 and Kimi K3's 2.94 across 203 tasks.
• Productivity-first data: Co-designed with CodeBuddy and WorkBuddy, so wins show up in real shipped work.

Getting Started

1. Download weights from tencent/Hy4-preview.
2. Deploy with vLLM or SGLang.
3. Call the OpenAI-compatible endpoint at http://127.0.0.1:8000/v1.
4. Start with a small task before scaling to full-context engineering work.

Frequently Asked Questions

Is the model open source?

Yes, Apache License 2.0. Weights are downloadable from multiple hubs.

What GPU do I need?

A single model of this scale (770B total) requires multi-GPU or a well-provisioned inference server. There is no hosted API at release.

Does it run a hosted API?

Tencent does not ship a hosted API for the preview; deploy vLLM or SGLang yourself.

Alternatives

• GLM-5.3: Zhipu's open front-end model, close competitor on the blind eval.
• DeepSeek V4: Another open-weight reasoning flagship.
• Kimi K3: Moonshot's open-weight model, close competitor on the blind eval.

Tips

1. Use a long context window deliberately; a 1M window is costly if you do not need it.
2. Prefer SGLang or vLLM with tensor parallelism for 49B activated params.
3. Treat the 203-task blind eval as a directional signal, not a universal ranking.