# Hy3

Hy3 activates 21B of its 295B parameters per token and exposes three reasoning levels behind a single `reasoning_effort` setting. Tencent Cloud aims Hy3 at agent work that needs dependable tool calls across long contexts.

- **Model ID:** `tencent/hy3`
- **Type:** chat
- **Providers:** deepinfra, novita, gmicloud, tencent
- **Context window:** 262,144
- **Maximum output tokens:** 262,144
- **Pricing:** $0.13/1M input tokens, $0.528/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/hy3

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'tencent/hy3',
  prompt: 'Why is the sky blue?'
})
```

## About

Tencent Cloud released Hy3 on July 6, 2026 under the Apache 2.0 license, publishing BF16 and FP8 weights on Hugging Face, ModelScope, GitCode, and CNB. Hy3 comes out of the Hunyuan line and follows the Hy3 Preview release from late April 2026.

The architecture is a Mixture-of-Experts (MoE) with 295B total parameters and 21B active per token, spread across 80 layers and 192 routed experts with top-8 routing. Attention is grouped-query with 64 heads and eight key-value heads. A separate 3.8B multi-token prediction layer drafts more than one token per forward pass, which serving stacks use for speculative decoding. The context window is 262.1K tokens, and one request can return up to 262.1K tokens.

Reasoning depth is a per-request setting. The `reasoning_effort` field accepts `no_think` for a direct answer, `low` for shallow reasoning, and `high` for deep chain-of-thought on math, coding, and analysis. `no_think` is the default, so you opt into reasoning tokens rather than paying for them on every call.

Tencent Cloud focused on tool-call and output-format stability, drawing on feedback from more than 50 of its own products. On SWE-Bench Verified, Hy3 holds accuracy variance within 4% across agent scaffoldings including CodeBuddy, Cline, and KiloCode, so a result from one harness carries to another. Internal evaluations built on real-world scenarios put the hallucination rate at 5.4%, down from 12.5%, and the commonsense error rate at 12.7%, down from 25.4%. On a multi-turn suite covering coreference resolution, ellipsis recovery, and constraint inheritance, the issue rate fell from 17.4% to 7.9%.

Tencent Cloud also ran a blind evaluation in which 270 experts scored tasks drawn from their own work. Hy3 averaged 2.67 out of four against GLM-5.1 at 2.51, with the widest margins on frontend development, data and storage, and CI/CD tasks.

Call Hy3 with one API key and get provider routing, automatic failover, and built-in observability. Integrate through the AI SDK, the Chat Completions API, the Responses API, the Messages API, or other supported API formats. Pay $0.13 per million input tokens, $0.528 per million output tokens, and $0.033 per million cached input tokens at current list rates.

## What to consider

Hy3 defaults to `no_think`, so a plain request returns a direct answer with no reasoning tokens. Set `reasoning_effort` to `high` on math, coding, and analysis work, and expect longer outputs and higher output-token spend when you do. Harness support for the field varies, so confirm your client forwards it.

Tencent Cloud recommends sampling at `temperature=0.9` and `top_p=1.0`. Start there before you tune, because the published quality figures assume those settings.

Hy3 takes text in and returns text out. Route requests that carry images, audio, or video to a multimodal model.

The hallucination, multi-turn, and expert-panel figures come from Tencent Cloud's own evaluations rather than a neutral third party. Validate on your own workload before you shift production traffic. For current throughput and latency, see live metrics on this page.

## When to use

### Best For

- Agent loops where 21B active parameters keep per-request compute low on tool-heavy runs
- Long-context work across repositories, contracts, and report sets that fill the 262.1K tokens window
- Mixed traffic that switches between direct answers and deep chain-of-thought per request
- Multi-turn assistants that must track intent, references, and constraints over long dialogues
- Grounded question answering where a fabricated detail costs more than a missing answer

### Consider Alternatives When

- Your inputs include images, audio, or video, since Hy3 handles text only
- Larger flagship models still lead on the hardest coding suites
- Your client cannot pass `reasoning_effort`, which leaves Hy3 on direct answers
- Short single-turn chat is the whole workload and reasoning depth adds no value

## Best for

- **Tool-Heavy Agent Loops:** Multi-step runs where 21B active parameters keep per-request compute low
- **Long-Context Analysis:** Repositories, contracts, and report sets that fill the 262.1K tokens window
- **Mixed Reasoning Traffic:** Requests that switch between direct answers and deep chain-of-thought
- **Multi-Turn Assistants:** Long dialogues that must track references, intent, and inherited constraints
- **Grounded Question Answering:** Work where a fabricated detail costs more than a missing answer

## Consider alternatives

- **Multimodal Inputs:** Hy3 handles text only, so send images, audio, or video to a vision-capable model
- **Frontier Coding Ceilings:** Larger flagship models still lead on the hardest coding suites
- **Fixed Client Parameters:** A client that cannot pass `reasoning_effort` leaves Hy3 on direct answers
- **Short Single-Turn Chat:** Reasoning depth adds output overhead that brief exchanges never recover

## Frequently asked questions

### What is Hy3?

Hy3 is a text-only Mixture-of-Experts language model from the Tencent Cloud Hunyuan team, released on July 6, 2026. It carries 295B total parameters with 21B active per token and succeeds the Hy3 Preview from late April 2026.

### What architecture does Hy3 use?

A Mixture-of-Experts (MoE) design with 295B total parameters, 21B active per token, 80 layers, and 192 routed experts using top-8 routing. Attention is grouped-query with 64 heads and eight key-value heads, and a 3.8B multi-token prediction layer supports speculative decoding.

### How do the three reasoning modes work?

Set `reasoning_effort` on the request. `no_think` returns a direct answer and is the default, `low` applies shallow reasoning, and `high` applies deep chain-of-thought for math, coding, and analysis. Providers expose the field differently, so check that your client forwards it, and remember reasoning tokens count toward your output budget.

### What is the context window for Hy3?

Hy3 supports a context window of 262.1K tokens and a max output of 262.1K tokens per request. That covers long agent transcripts, large document sets, and repository-scale code review.

### Is Hy3 open source?

Yes. Tencent Cloud released Hy3 under the Apache License 2.0, with BF16 and FP8 weights on Hugging Face, ModelScope, GitCode, and CNB. Through AI Gateway you use hosted inference, so licensing matters only if you also run the weights yourself.

### What sampling parameters does Tencent Cloud recommend?

`temperature=0.9` and `top_p=1.0`. Published quality figures assume those settings, so treat them as the starting point before you tune for your own task.

### How consistent is Hy3 across agent frameworks?

SWE-Bench Verified accuracy varies by less than 4% across scaffoldings including CodeBuddy, Cline, and KiloCode. That narrow spread means a result you measure in one harness generally holds when you switch to another.

### How much does Hy3 hallucinate?

Tencent Cloud's internal evaluations on real-world scenarios report a 5.4% hallucination rate, down from 12.5%, and a 12.7% commonsense error rate, down from 25.4%. The stated design target is to answer when grounded and say so when evidence is missing.

### How do I call Hy3 through AI Gateway?

Use the model identifier `tencent/hy3` with the AI SDK, the Chat Completions API, the Responses API, the Messages API, or another supported API format. You authenticate with an AI Gateway API key, and no Tencent Cloud account is needed.

### Does AI Gateway support Zero Data Retention for Hy3?

Yes, Zero Data Retention is available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

## Links

- [Model page](https://vercel.com/ai-gateway/models/hy3)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://deepinfra.com/tencent/Hy3)