Hy3-preview
Hy3-preview is a preview-stage large language model from Tencent, currently tracked on Artificial Analysis throughput benchmarks at roughly 74 output tokens per second.
API Pricing
Cheapest on OpenRouter — 4% below avg| Provider | Input / 1M | Output / 1M | Cached / 1M |
|---|---|---|---|
| $0.132 | $0.528 | $0.033 | |
| $0.140 | $0.580 | $0.035 | |
| $0.140 | $0.580 | $0.035 |
Prices updated daily. Last check: Oct 9, 2026
Hy3-preview pricing by provider
Input, output, and batch rates, plus alternatives, for one provider at a time.
Performance & Benchmarks
Source: Artificial Analysis →Reasoning & Knowledge
- GPQA Diamond89.7%
- Humanity's Last Exam33.5%
Coding
- SciCode48.6%
Agentic & Tool Use
- Terminal-Bench v2.164.4%
- τ-bench Banking22.9%
Instruction & Long Context
- Long-Context Reasoning79.0%
Benchmarks measured Oct 2026. Scores are independent evaluations, not vendor-reported.
Model Details
General
- Creator
- Tencent
- Modalities
- Text
Capabilities
- Open Source
- Yes
Strengths & Limitations
Strengths
- Measured generation throughput of roughly 73.6 output tokens per second in Artificial Analysis testing, giving a concrete performance data point for planning
- Preview access allows teams to evaluate a newer Tencent model generation before committing to a production dependency
- Backed by Tencent, which maintains an ongoing line of in-house language models rather than a single one-off release
- Independent third-party latency and throughput measurements are available, rather than only vendor-published figures
- Useful as a comparison baseline when benchmarking against other providers' models on the same prompts
Limitations
- Preview status means behavior, availability, and endpoints can change without the stability guarantees of a general-availability release
- Time to first token of roughly 1,787 ms is high enough to be noticeable in interactive chat or voice interfaces
- We do not currently track a confirmed context window for this model, so long-document limits should be verified with the serving provider
- No verified accuracy benchmark scores are available in our database, making direct quality comparisons difficult
- Provider availability may be narrower than for widely distributed models, which limits routing and failover options
Key Features
About Hy3-preview
Common Use Cases
Hy3-preview is best suited to evaluation and comparison workloads rather than latency-critical production paths. Its measured throughput of roughly 74 output tokens per second supports batch and asynchronous jobs — content drafting, summarization runs, offline data processing, and prompt regression testing — where a per-request first-token delay near 1.8 seconds is absorbed by the pipeline rather than felt by a user. Teams already evaluating models across multiple vendors can use it as an additional comparison point when assessing Tencent's model line. Because it is a preview build and we do not have a confirmed context window or accuracy benchmark record for it, workloads that depend on very long inputs or on documented reasoning quality should be validated directly against the provider's published specs before rollout.
Frequently Asked Questions
How much does Hy3-preview cost to use?
Pricing varies by provider and by pricing type (input tokens, output tokens, cached input, and any batch discounts). Because rates change frequently, we do not quote figures in this write-up — see the pricing table on this page for current per-provider rates.
What is Hy3-preview best used for?
It fits evaluation work and asynchronous or batch generation tasks such as summarization, drafting, and offline processing. Its measured time to first token of around 1,787 ms makes it less well suited to highly interactive chat or voice experiences where initial response delay is felt directly.
How fast is Hy3-preview?
Artificial Analysis measurements in our database show approximately 73.6 output tokens per second of generation throughput and about 1,787 ms time to first token. Actual figures depend on the serving provider, prompt length, and load at the time of the request.
What does the "preview" in Hy3-preview mean?
It indicates a pre-general-availability build. Preview models are typically offered so developers can test capabilities early, but their behavior, availability, and API details can change between iterations, so pinning a production workload to one carries migration risk.
What context window does Hy3-preview support?
We do not currently have a verified context window for this model in our database. Check the specifications published by the provider you plan to route through, since context limits can also differ between hosts of the same model.
Who created Hy3-preview?
Hy3-preview comes from Tencent, which develops and releases its own family of large language models.