Overview
Tencent Hy3 Preview is the first major model released under the new Tencent HY brand on April 23, 2026 — and the first flagship model trained after Tencent's February 2026 rebuild of its pre-training and reinforcement-learning infrastructure, led by chief AI scientist Yao Shunyu. Multiple reports state that Hy3 Preview began training in late January 2026 and reached public release in under three months on the rebuilt stack — keep this framed as the reported timeline rather than an independently verified official changelog fact.
Hy3 Preview is a 295-billion-parameter Mixture-of-Experts language model with 21 billion active parameters, an additional 3.8B Multi-Token Prediction (MTP) layer, and a 256K-token context window. Tencent released model weights on Hugging Face, ModelScope, and GitCode under the custom Tencent Hy Community License Agreement, with the Tencent-Hunyuan GitHub repository hosting the project code, documentation, training pipeline, and license. Hosted access is available through Tencent Cloud TokenHub with tiered pricing starting at CNY 1.2/M input and CNY 4/M output for the 0–16K tier, plus a currently free Hy3 Preview route listed on OpenRouter.
What's New
Fast-and-Slow-Thinking Fused Architecture
Hy3 Preview ships a unified architecture that switches between rapid responses and deeper deliberate reasoning per query, rather than forcing users to manually toggle between a "fast" base model and a separate "thinking" reasoning variant — the path the previous Hunyuan generation took with TurboS as the fast base and the in-depth Hunyuan-T1 layered on top. The fused approach simplifies deployment for agent workflows that mix short tool calls with multi-step planning in a single conversation.
295B / 21B Active MoE with Sparse Expert Activation
The architecture uses a dense-MoE hybrid: 192 routed experts plus one always-active shared expert per MoE layer, across 80 transformer layers with 64 grouped-query attention heads. Sparse expert activation is designed to improve serving efficiency, with 21B active parameters per token versus 295B total. A 3.8B Multi-Token Prediction layer accelerates token generation without changing the underlying expert routing.
Compared with Hunyuan 2.0 (406B total / 32B active), Hy3 Preview is a smaller and more sparsely activated flagship — Tencent's framing positions it as the first deliverable from the rebuilt training stack rather than as a strict size upgrade.
Strong Coding and Agent Benchmark Results
Hy3 Preview targets reasoning, instruction following, coding, and agent tasks. Tencent's published numbers at launch:
- SWE-bench Verified: 74.4
- Terminal-Bench 2.0: 54.4
- BrowseComp: 67.1
- WideSearch: 70.2
These place Hy3 Preview within range of Claude Opus 4.6 (80.8 on SWE-bench Verified) and below GLM-5 (77.8) and Kimi K2.5 (76.8) on the same evaluation, while still being competitive given the substantially lower API cost.
Open-Weight Release Under Tencent Hy Community License
Model weights are distributed through Hugging Face (tencent/Hy3-preview), ModelScope, and GitCode, while GitHub (Tencent-Hunyuan/Hy3-preview) hosts the repository, documentation, training assets, and license. The license includes territory restrictions, a >100M monthly active user commercial-license trigger, and limits on using model outputs to improve other AI models. The release includes a complete training pipeline supporting both full fine-tuning and LoRA fine-tuning, with DeepSpeed ZeRO configurations and LLaMA-Factory integration.
First Model Under the Tencent HY Brand
Hunyuan was officially rebranded to Tencent HY in late 2025 to support clearer global positioning. Hy3 Preview is the first major model released under the new brand convention; earlier models continue to be referenced under their original Hunyuan names.
Performance Benchmarks
| Benchmark | Hy3 Preview | Claude Opus 4.6 | GLM-5 | Kimi K2.5 |
|---|---|---|---|---|
| SWE-bench Verified (coding agent) | 74.4 | 80.8 | 77.8 | 76.8 |
| Terminal-Bench 2.0 (agent shell tasks) | 54.4 | — | — | — |
| BrowseComp (search agent) | 67.1 | — | — | — |
| WideSearch (search agent breadth) | 70.2 | — | — | — |
| Total parameters | 295B (MoE) | Undisclosed | Undisclosed | 1T (MoE) |
| Active parameters | 21B | Undisclosed | Undisclosed | ~32B |
| Context window | 256K | 1M | 1M | 256K |
| Open weights | Yes (custom license) | No | Partial | Partial |
| Lowest reported input price | Limited-time free routes where available, then CNY 1.2/M input ($0.17/M, FX-dependent) for 0–16K on Tencent Cloud TokenHub | $5/M | Varies | Varies |
Hy3 Preview's positioning is clearer when paired with the cost column: its published SWE-bench Verified score trails the listed competitors while TokenHub pricing starts substantially lower than closed frontier API pricing, with open weights for self-hosting.
Pricing & Plans
Hy3 Preview uses a freemium/open-weight access model: model weights are available for self-hosting, OpenRouter currently lists a free hosted route, and Tencent Cloud TokenHub provides paid API access with tiered token pricing.
Free Access
- Hy3 Preview release page at hy.tencent.com; the official repository links Tencent AI Studio as the product site
- Limited-time free hosted testing announced via OpenRouter
- Open-weight access via Hugging Face, ModelScope, and GitCode for self-hosting, plus the Tencent-Hunyuan GitHub repository for deployment and training documentation
Tencent Cloud TokenHub API (tiered by input length)
| Input length tier | Input | Output | Cache hit |
|---|---|---|---|
| 0–16K tokens | CNY 1.2 / M tokens (~$0.17/M, FX-dependent) | CNY 4 / M tokens (~$0.55/M, FX-dependent) | CNY 0.4 / M tokens |
| 16K–32K tokens | CNY 1.6 / M tokens (~$0.22/M, FX-dependent) | CNY 6.4 / M tokens (~$0.88/M, FX-dependent) | CNY 0.6 / M tokens |
| 32K+ tokens | CNY 2 / M tokens (~$0.28/M, FX-dependent) | CNY 8 / M tokens (~$1.10/M, FX-dependent) | CNY 0.8 / M tokens |
Pricing is set by Tencent Cloud TokenHub. Verify current rates, regional availability, and any active limited-time free routes on Tencent Cloud and OpenRouter before production deployment.
Best For
- Cost-conscious teams building coding agents, search agents, and reasoning workloads who can tolerate "preview" stability for substantially lower per-token cost
- ML engineers and researchers who want an open-weight 295B MoE reasoning model with a published training pipeline for fine-tuning experiments
- Developers comparing Chinese frontier reasoning models — Hy3 Preview, GLM-5, Kimi K2.5 — to find the best fit for agentic workflows
- Enterprises in supported Tencent Cloud regions standardizing on Tencent Cloud, after confirming license territory, API availability, and compliance requirements
FAQ
What is Hy3 Preview?
Hy3 Preview is Tencent's April 23, 2026 release of a 295B-parameter Mixture-of-Experts language model with 21B active parameters, a 3.8B MTP layer, and 256K context. It is the first major model under the Tencent HY brand and the first flagship trained on Tencent's rebuilt pre-training and reinforcement-learning stack.
How does Hy3 Preview compare to the previous Hunyuan flagship?
Hunyuan 2.0 (December 2025) used roughly 406B total parameters with 32B active. Hy3 Preview is smaller and more sparsely activated (295B / 21B active), framed by Tencent as the first deliverable from the rebuilt training infrastructure rather than a pure size upgrade. The architecture also fuses fast and slow thinking into a single model rather than splitting them across TurboS and Hunyuan-T1 as the previous generation did.
Is Hy3 Preview really open source?
Hy3 Preview is open-weight rather than OSI-standard open source. Weights are available on Hugging Face, ModelScope, and GitCode under the custom Tencent Hy Community License Agreement; the Tencent-Hunyuan GitHub repository provides project code, deployment guidance, training documentation, and license files. The license includes territory restrictions, a >100M MAU commercial-license trigger, and limits on using model outputs to improve other AI models.
How can I try Hy3 Preview for free?
Tencent has announced a limited-time free hosted period on Tencent Cloud and limited-time free hosted access via OpenRouter. The Hy3 Preview release page is at hy.tencent.com, and the official repository links Tencent AI Studio as the product site. You can also download weights from Hugging Face at tencent/Hy3-preview for self-hosting.
What does Hy3 Preview cost on Tencent Cloud after the free period?
Tencent Cloud CN / TokenHub lists tiered pricing: CNY 1.2/M input and CNY 4/M output for 0–16K inputs, CNY 1.6/M input and CNY 6.4/M output for 16K–32K inputs, and CNY 2/M input and CNY 8/M output for 32K+ inputs (roughly $0.17–$0.28/M input, FX-dependent). Confirm current rates and regional availability directly on Tencent Cloud before production use.
How does Hy3 Preview compare to Claude Opus 4.6 on coding?
On SWE-bench Verified, Hy3 Preview scores 74.4 versus Claude Opus 4.6's 80.8 — a 6-point gap. The trade-off is API cost: Hy3 Preview's post-free Tencent Cloud price is roughly $0.28/M input versus $5/M for Claude Opus 4.6, and Hy3 Preview is open-weight while Claude is closed.
