Hy3-preview (Non-reasoning)
Tencent • text#72 of 571 in intelligence
Hy3 is a 295B-parameter Mixture-of-Experts model from Tencent (21B active, 192 experts with top-8 routing) built for reasoning, agentic workflows, and real-world production use. It supports a configurable reasoning effort:...
Specifications
Context Window
262K tokens
Input Price/1M
$0.06
#224 of 686
Output Price/1M
$0.21
Parameters
—
Speed
72 tok/s
#176 of 525
Latency (TTFT)
2.8s
Max Output
128K tokens
Cost per Task
433
AA Score / US$ · higher = better
Benchmarks
Hy3-preview (Non-reasoning) results on the main AI model evaluation benchmarks. Higher scores indicate better performance.
Coding
| Benchmark | Score | Maximum | Methodology |
|---|---|---|---|
| AA Coding Index | 58.8 | 100.0 | — |
overall
| Benchmark | Score | Maximum | Methodology |
|---|---|---|---|
| AA Intelligence Index | 42.2 | 100.0 | Artificial Analysis official API |
Information
- Release date
- April 23, 2026
- Tool Calling
- ✅ Supported
- Vision
- ❌ Not supported
- Audio
- ❌ Not supported
Full Analysis: Hy3-preview (Non-reasoning)
What is Hy3-preview (Non-reasoning)?
Hy3-preview (Non-reasoning) is an AI model developed by Tencent, classified as a text model. It focuses on text processing and natural language generation. As a proprietary model, it is available via Tencent's cloud API. With a context window of 262K tokens, it is suitable for processing long documents such as contracts, books, and complete codebases.
Pricing & Costs in 2026
Hy3-preview (Non-reasoning) is usage-based, priced at $0.06/1M input tokens and $0.21/1M output tokens. For context: 1 million tokens is approximately 750,000 words, or about 10 average-length books. At this aggressive price point, it is one of the most cost-effective options on the market, ideal for high-volume applications like chatbots, bulk document analysis, and automation.
Benchmarks & Performance
Hy3-preview (Non-reasoning) was evaluated on 2 different benchmarks, covering categories like Coding, overall. Results show moderate performance across available evaluations.
It's important to note that benchmarks measure specific aspects and don't capture the full user experience. Factors like instruction adherence, behavior in long conversations, and real-world task quality vary significantly between models and aren't always reflected in standard scores.
Recommended Use Cases
Hy3-preview (Non-reasoning) specializes in text, offering advanced capabilities for creating and processing text content.
Comparison with Alternatives
In the 2026 AI model ecosystem, Hy3-preview (Non-reasoning) competes directly with similarly capable models. Tencent competes in this segment against OpenAI, Anthropic, Google, and Meta. The choice between models depends on the specific use case, budget, latency requirements, and need for features like multimodality and tool calling.
For a detailed side-by-side comparison, use our comparison tool or check the overall model ranking.
Frequently Asked Questions
What is Hy3-preview (Non-reasoning)?
Hy3-preview (Non-reasoning) is an AI model developed by Tencent. It is a text model.
How much does Hy3-preview (Non-reasoning) cost?
Hy3-preview (Non-reasoning) costs $0.06/1M input tokens and $0.21/1M output tokens. For heavy usage (e.g., a chatbot handling 100k messages/month), costs can range from $10 to $1,000 depending on volume.
How does Hy3-preview (Non-reasoning) compare with other models?
In available benchmarks, Hy3-preview (Non-reasoning) scored: AA Coding Index: 58.8/100, AA Intelligence Index: 42.2/100. See the full table above for a detailed comparison.
Is Hy3-preview (Non-reasoning) open source?
No, Hy3-preview (Non-reasoning) is a proprietary model from Tencent. It is available via cloud API. For open source alternatives, check our open source model ranking.
What is Hy3-preview (Non-reasoning) best for?
Hy3-preview (Non-reasoning) excels at general-purpose language tasks. With its large context window, it handles long documents, codebases, and extended conversations. It supports tool calling for API integrations and automation.
Last updated: August 18, 2026 • View methodology →