Hy3: Complete Specifications, Pricing, API Access & Use Cases (2026)

Hy3 is Tencent’s production reasoning and agent model designed for complex planning, coding, document-intensive tasks and tool-driven workflows. Released on July 6, 2026, it succeeds Hy3 Preview with additional post-training, improved stability and stronger agent behavior. Tencent documents 295 billion total parameters, 21 billion active parameters and a 256K context window, while the supplied Gate.AI listing reports a 262K platform context value.

What Is Hy3?

Hy3 is a Mixture-of-Experts (MoE) large language model developed by the Tencent Hy Team. It uses 295B total parameters but activates approximately 21B for each token, alongside a 3.8B-parameter multi-token-prediction layer. Tencent describes it as a hybrid fast-and-slow-thinking model built for reasoning, agents and productivity workloads.

The production release followed Hy3 Preview, which launched in April 2026. Tencent says feedback from more than 50 products informed expanded post-training before the final Hy3 release. This distinction matters for search intent: Hy3 Preview covers the earlier experimental version, while this page focuses specifically on the July production model, its revised pricing, configurable reasoning and deployment profile.

Tencent has also released Hy3 weights under Apache 2.0 and provides deployment guidance for vLLM and SGLang.

What Are Hy3’s Key Specifications and Pricing?

Tencent’s official repository specifies a 256K context length, not 262K. The 262K figure comes from the Gate.AI model-card details supplied for this article and should therefore be treated as the platform-specific hosted limit rather than silently replacing Tencent’s model specification.

A practical Gate.AI cost example shows why the model can be attractive for long agent runs. At the listed rates, a request using 500,000 uncached input tokens and producing 100,000 output tokens would cost approximately:

Input: 0.5 × $0.132 = $0.066
Output: 0.1 × $0.528 = $0.0528
Estimated total: $0.1188

This is a calculated example before retries, additional agent turns or account-specific billing adjustments.

What Can Hy3 Do That Makes It Useful in Production?

Hy3 is particularly relevant to long-horizon agents. Tencent reports improvements to tool-call stability, output-format consistency and resistance to intent drift across extended interactions. Those characteristics matter when an agent must repeatedly interpret intermediate results rather than simply answer one prompt.

Its configurable reasoning is another useful distinction. Tencent’s open-source deployment supports no_think, low and high reasoning-effort settings, allowing applications to use faster direct responses for simple work while allocating deeper reasoning to coding, mathematics or complex planning.

For coding and technical agents, Hy3 can support repository analysis, debugging plans, CI/CD reasoning and multi-stage implementation tasks. Tencent specifically highlights improvements across frontend development, data and storage work, financial modeling and productivity workflows. These are provider-reported capabilities and should still be validated against a team’s own workloads.

The model’s long context also makes it suitable for large documents, financial statements, specifications or collections of code files where aggressive chunking would otherwise lose cross-document relationships.

What Are Hy3’s Supported Modalities?

Hy3 should therefore be treated primarily as a text and code reasoning model, not as a general native multimodal system. Tencent’s wider Hunyuan portfolio includes image, video and other multimodal technologies, but those capabilities should not be attributed to Hy3 without model-specific documentation.

Where Does Hy3 Fall Short?

The biggest practical limitation is that a 256K or platform-listed 262K context window is still finite. Very large repositories, document collections and persistent agent histories may require retrieval, summarization or context pruning.

A second consideration is self-hosting infrastructure. Tencent recommends eight high-memory GPUs such as H20-class hardware for deployment guidance. The model activates only 21B parameters at a time, but its 295B total parameter footprint still makes local production serving substantially more demanding than deploying a small dense model.

Some hosted-service details are also not publicly confirmed for the exact Gate.AI route, including Hy3-specific rate limits, maximum output length, cache-write pricing and whether Tencent’s reasoning_effort control is exposed identically through Gate.AI.

Finally, improved reliability does not make model output automatically correct. Code, financial analysis, document extraction and autonomous tool actions should still use validation, permissions and human-review controls.

What Is Hy3 Best Used For?

Hy3 fits multi-step agents that must plan, call tools and preserve goals across extended sessions. Its reasoning-effort controls also make it useful when an application mixes simple requests with harder tasks.

It can suit coding and repository workflows where the model must understand several files, reason about dependencies and propose coordinated changes. Teams focused specifically on autonomous coding can also compare GLM 5.1, which is positioned around long-horizon software engineering.

Hy3 can also support long-document analysis and financial modeling where significant source material must remain available during reasoning. Generated financial conclusions should be independently verified rather than treated as professional financial advice.

Choose Hy3 when low listed token pricing, long context, configurable reasoning and agent-oriented behavior are priorities. Consider another model when native vision/audio support, extremely large context beyond 256K-class limits or lighter self-hosting requirements are more important.

How Does Hy3 Compare to Hy3 Preview and GLM 5.1?

Hy3 is the logical replacement for Hy3 Preview where the final production model is available. Tencent says the newer model improves stability, cost efficiency and agent capability, so the Preview article should remain focused on the April release rather than competing for the same primary query.

Against GLM 5.1, Hy3 has substantially lower listed Gate.AI token pricing and a larger documented model context, while GLM 5.1 has a particularly explicit long-horizon coding-agent positioning. The appropriate choice depends on tool reliability, task success rate, latency and total workflow cost rather than token price alone.

How Do I Access Hy3 Through Gate.AI?

Gate.AI’s documentation verifies an OpenAI-compatible chat-completions route at https://api.gate.ai/openai/v1/chat/completions, using Bearer authentication and the model identifier in the request body. Its model-list API provides platform model IDs, context lengths and pricing.

For Hy3, use the Gate.AI model ID:

tencent/hy3

Python Example

import os
from openai import OpenAI
client = OpenAI(
    api_key=os.environ["GATEAI_API_KEY"],
    base_url="https://api.gate.ai/openai/v1",
)
try:
    response = client.chat.completions.create(
        model="tencent/hy3",
        messages=[
            {
                "role": "user",
                "content": "Create a step-by-step migration plan for this Python service."
            }
        ],
    )
    print(response.choices[0].message.content)
except Exception as exc:
    print(f"Request failed: {exc}")

curl Example

curl https://api.gate.ai/openai/v1/chat/completions \
  -H "Authorization: Bearer $GATEAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "tencent/hy3",
    "messages": [
      {
        "role": "user",
        "content": "Identify the main risks in this deployment plan."
      }
    ]
  }'

These examples use Gate.AI’s documented OpenAI-compatible request structure. They intentionally omit Hy3-specific reasoning controls because the available Gate.AI documentation does not confirm that Tencent’s self-hosted reasoning_effort parameter is exposed through the hosted route. Gate.AI’s SDK documentation confirms chat-completions support, while applications remain responsible for agent loops, tool dispatch, memory and orchestration.

FAQs

Is Hy3 the same as Hy3 Preview?

No. Hy3 Preview was released in April 2026. The production Hy3 model followed on July 6, with additional post-training and provider-reported improvements in stability, cost efficiency and agent performance.

What is Hy3’s context window?

Tencent officially documents a 256K-token context window. The supplied Gate.AI model card lists 262K, so the two values should be treated as provider-model and hosted-platform specifications respectively rather than assumed to be identical.

How much does Hy3 cost on Gate.AI?

The supplied Gate.AI record lists $0.132 per 1M input tokens, $0.528 per 1M output tokens and $0.033 per 1M cache-read tokens. Cache-write pricing is not listed.

Does Hy3 support configurable reasoning?

Yes at the Tencent model/deployment level. Tencent documents no_think, low and high reasoning-effort modes. Whether the same control is exposed identically through every hosted provider should be verified before implementation.

Is Hy3 open source?

Tencent publishes Hy3 weights and deployment resources under the Apache License 2.0, including guidance for vLLM and SGLang.

What is the correct Gate.AI model ID for Hy3?

For this Hy3 page, the model identifier is tencent/hy3. The bytedance/seed-2-1-turbo identifier belongs to a different ByteDance model and should not be used in Hy3 API requests.