Hy3 Preview: Complete Specifications, Pricing, API Access & Use Cases (2026)
What Is Hy3 Preview?
Hy3 Preview is Tencent’s 295B-parameter Mixture-of-Experts reasoning and agent model, released in April 2026, featuring a 256K-token context window and text-generation capabilities, with Gate.AI pricing listed at \$0.18 per 1M input tokens and \$0.60 per 1M output tokens as of July 2026.
Tencent describes Hy3 Preview as a MoE model developed by the Tencent Hy Team with 295B total parameters, 21B active parameters, and 3.8B MTP layer parameters. Tencent’s official materials position the model for complex reasoning, instruction following, context learning, coding, and agent tasks.
Hy3 Preview is relevant for developers and AI teams evaluating long-context reasoning models for agent workflows, code assistance, document processing, and multi-step planning. Tencent states that Hy3 Preview is open-sourced and available through GitHub, Hugging Face, ModelScope, and GitCode, with mainstream inference-framework support including vLLM and SGLang.
What Are Hy3 Preview’s Key Specifications and Pricing?
| Field | Value |
|---|---|
| Provider | Tencent / Tencent Hy Team (as of July 2026) |
| Model Family | Hy / Tencent Hy model family (as of July 2026) |
| Model Type | Mixture-of-Experts reasoning and agent model; text-generation model (as of July 2026) |
| Release Date | April 2026; Tencent announcement confirms Hy3 Preview availability and open-source release (as of July 2026) |
| Context Window | 256K tokens (as of July 2026) |
| Total Parameters | 295B parameters (as of July 2026) |
| Activated Parameters | 21B active parameters (as of July 2026) |
| MTP Layer Parameters | 3.8B MTP layer parameters (as of July 2026) |
| Architecture | Mixture-of-Experts, 192 experts, top-8 activated, BF16 precision listed in model card (as of July 2026) |
| Input Pricing | \$0.18 per 1M input tokens, as per Gate.AI listing and aligned with Tencent Cloud starting input pricing of approximately \$0.18 per 1M tokens (as of July 2026) |
| Cached Input Pricing | \$0.06 per 1M cache-read tokens, as per Gate.AI listing and aligned with Tencent Cloud cached input pricing of approximately \$0.06 per 1M tokens (as of July 2026) |
| Cache Write Pricing | Not specified in the availableGate.AIlisting as of July 2026 |
| Output Pricing | \$0.60 per 1M output tokens, as per Gate.AI listing; Tencent Cloud states output starts at approximately \$0.59 per 1M tokens (as of July 2026) |
| Pricing Unit | Per 1M tokens (as of July 2026) |
| Modality Support | Text input and text output verified; image, audio, and video support not confirmed for Hy3 Preview as of July 2026 |
| Supported Input Types | Text prompts, long-context text inputs, code, and structured natural-language instructions (as of July 2026) |
| Supported Output Types | Text-generation responses (as of July 2026) |
| API Access | Gate.AI OpenAI-compatible API access; Tencent Cloud TokenHub availability; open-weight deployment via GitHub, Hugging Face, ModelScope, and GitCode (as of July 2026) |
| Model ID | tencent/hy3-preview, as perGate.AIlisting (as of July 2026) |
| Availability | Gate.AI model catalog listing, Tencent Cloud TokenHub, OpenRouter listing noted by Tencent, and open-weight repositories (as of July 2026) |
| Knowledge Cutoff | Not confirmed from official sources as of July 2026 |
| Rate Limits | Account-level and route-level limits may apply; exact Hy3 Preview limits not confirmed from official sources as of July 2026 |
| Fine-tuning Support | Tencent repository references training and LoRA/full fine-tuning resources; managed Gate.AI fine-tuning support not confirmed as of July 2026 |
| Streaming Support | Gate.AI General Chat API lists chat completions with streaming support; Hy3-specific streaming behavior should be validated per deployment (as of July 2026) |
| Batch API Support | Not confirmed from official sources as of July 2026 |
| Tool / Function Calling | Tencent describes agent-framework integration and toolchain orchestration use cases; exact managed tool-calling schema for Hy3 Preview should be validated before production use (as of July 2026) |
| Structured Output / JSON Mode | Not confirmed from official sources as of July 2026 |
| License / Usage Restrictions | Tencent Hy Community License Agreement (as of July 2026) |
What Can Hy3 Preview Do That Makes It Useful in Production?
Hy3 Preview is useful for long-context reasoning workflows because Tencent’s official model materials list a 256K-token context length. This can support large prompts, long documents, multi-file analysis, and agent memory blocks when applications include retrieval checks and token-budget controls. Understanding AI model context windows helps teams estimate how much material a model can process in one request.
Hy3 Preview is also relevant for agent workflows. Tencent reports that the model has been integrated across products including Yuanbao, ima, CodeBuddy, WorkBuddy, QQ, QQ Browser, Tencent Docs, and Tencent LearnShare, and states that it has supported complex real-user agent workflows of up to 495 steps. These claims are provider-reported and should be validated on each team’s own workload before production adoption.
For coding and software-agent experiments, Tencent highlights improved code and agent capabilities, while the official repository includes deployment guidance for vLLM and SGLang. This makes Hy3 Preview relevant for repository analysis, code explanation, planning, debugging support, and multi-step engineering assistants, provided generated code is reviewed and tested.
Hy3 Preview may also fit cost-aware reasoning deployments. As per the Gate.AI listing, the model’s token pricing is \$0.18 per 1M input tokens, \$0.06 per 1M cache-read tokens, and \$0.60 per 1M output tokens as of July 2026. Teams comparing it with DeepSeek-R1 reasoning workflows or Kimi K2 agent-oriented use cases should evaluate task reliability, latency, output length, tool behavior, and total workflow cost rather than token price alone.
What Are Hy3 Preview’s Supported Modalities?
| Modality | Supported? | Notes |
|---|---|---|
| Text input | Yes | Official examples and model cards describe text-generation usage, long-context prompts, code, and natural-language instructions (as of July 2026). |
| Text output | Yes | Hy3 Preview is listed as a text-generation model on Hugging Face and in Tencent materials (as of July 2026). |
| Image input | Not confirmed | No official Hy3 Preview image-input support was confirmed as of July 2026. |
| Audio input | Not confirmed | No official Hy3 Preview audio-input support was confirmed as of July 2026. |
| Video input | Not confirmed | No official Hy3 Preview video-input support was confirmed as of July 2026. |
| Embeddings | Not confirmed | Hy3 Preview is not documented as an embedding model as of July 2026. |
| Tool outputs | Partially confirmed conceptually | Tencent describes agent-framework and toolchain orchestration scenarios; exact managed API schema should be validated before use. |
Where Does Hy3 Preview Fall Short?
Hy3 Preview’s most important limitation is that several operational details are not fully specified in public official sources as of July 2026. Knowledge cutoff, model-specific managed rate limits, batch API support, structured output mode, and exact service-level terms should be verified before production deployment.
Self-hosting may require substantial infrastructure. Tencent lists Hy3 Preview as a 295B-parameter MoE model with 21B active parameters and BF16 precision, and provides deployment instructions for high-end inference frameworks. Even with MoE efficiency, local serving can require specialized GPUs, inference optimization, monitoring, and operational expertise.
This is a general AI limitation and is not model-specific unless stated by Tencent: reasoning models can generate plausible but incorrect outputs. Hy3 Preview responses should be checked against authoritative sources before they are used in legal, medical, financial, cybersecurity, safety-critical, or other high-stakes contexts.
High-reasoning workflows may also increase latency and output length. Teams should define prompt policies, output limits, retrieval constraints, logging rules, and human-review thresholds before using Hy3 Preview in customer-facing agents or automated decision workflows.
What Is Hy3 Preview Best Used For?
| Use Case | Why Hy3 Preview May Fit | Important Limitation |
|---|---|---|
| Multi-step agent workflows | Tencent positions Hy3 Preview for reasoning, coding, and agent tasks, and reports real-world agent workflow use across Tencent products (as of July 2026). | Agent behavior should be tested with real tools, permissions, and rollback controls. |
| Long-document analysis | The 256K-token context length can support long prompts and document sets (as of July 2026). | Long context does not guarantee perfect retrieval, citation accuracy, or instruction adherence. |
| Coding assistants | Tencent highlights code and agent gains and provides deployment examples (as of July 2026). | Generated code requires tests, security review, and developer supervision. |
| Complex planning | Hy3 Preview is positioned for complex reasoning and multi-step execution (as of July 2026). | Higher reasoning depth can increase latency and cost. |
| Enterprise experimentation with open weights | Tencent states the model is open-sourced on GitHub, Hugging Face, ModelScope, and GitCode (as of July 2026). | License terms, deployment costs, and governance requirements must be reviewed. |
| Cost-aware API routing | Gate.AIpricing is listed at low per-token rates for input, cache read, and output as of July 2026. | Total spend depends on prompt size, output length, retries, caching, and routing policy. |
How Does Hy3 Preview Compare to DeepSeek-R1 and Kimi K2?
| Comparison Area | Hy3 Preview | DeepSeek-R1 | Kimi K2 | Scenario Fit |
|---|---|---|---|---|
| Primary Positioning | Tencent MoE reasoning and agent model with 295B total parameters and 21B active parameters (as of July 2026). | Reasoning-focused model family often used for deliberate problem solving. | Agentic and coding-oriented model family often evaluated for tool workflows. | Hy3 Preview may fit teams prioritizing Tencent’s open-weight MoE design and long-context reasoning. |
| Context Window | 256K tokens verified from Tencent model materials (as of July 2026). | Deployment-specific context should be checked for the route used. | Deployment-specific context should be checked for the route used. | Hy3 Preview is attractive when a verified 256K context window is required. |
| Access Pattern | Gate.AIlisting, Tencent Cloud TokenHub, OpenRouter listing noted by Tencent, and open-weight repositories (as of July 2026). | Access varies by provider, gateway, and deployment route. | Access varies by provider, gateway, and deployment route. | Use route-specific evaluation before switching production workloads. |
| Agent Use | Tencent reports product integrations and complex agent workflow support (as of July 2026). | Often considered for reasoning chains and problem solving. | Often considered for coding and agent tasks. | Choose based on private task benchmarks, not broad "winner" claims. |
| Cost Evaluation | Gate.AIlisting: \$0.18 input, \$0.06 cache read, \$0.60 output per 1M tokens (as of July 2026). | Pricing depends on route and provider. | Pricing depends on route and provider. | Compare total workflow cost, including retries, cache hit rate, output length, and latency. |
How Do I Access Hy3 Preview Through Gate.AI?
Gate.AI lists Hy3 Preview in its model catalog and provides OpenAI-compatible API access through https://api.gate.ai/openai/v1. Gate.AI documentation states that the platform is fully compatible with the OpenAI API, supports Python, Node.js, curl, and common developer tools, and uses a Gate.AI API key with OpenAI-compatible requests.
As per the Gate.AI listing, the model ID is:
tencent/hy3-preview
Gate.AI documentation confirms the general chat API base URL, bearer-token authentication, OpenAI-compatible format, pay-as-you-go pricing, and the POST /chat/completions endpoint.
Python Example
from openai import OpenAIimport osclient = OpenAI(api_key=os.environ["GATEAI_API_KEY"],base_url="https://api.gate.ai/openai/v1",)completion = client.chat.completions.create(model="tencent/hy3-preview",messages=[{"role": "system","content": "You are a concise technical assistant."},{"role": "user","content": "Summarize when a high-reasoning model is useful for an agent workflow."}],)print(completion.choices[0].message.content)
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-preview","messages": [{"role": "system","content": "You are a concise technical assistant."},{"role": "user","content": "Explain when Hy3 Preview may fit long-context agent workflows."}]}'
Through Gate.AI, developers can use a unified OpenAI-compatible request pattern for supported models. Before production deployment, teams should verify account-level model availability, routing settings, rate limits, billing behavior, caching behavior, and any Hy3-specific request parameters in the Gate.AI console.
FAQs
What is Hy3 Preview’s context window?
Hy3 Preview’s official Tencent model materials list a 256K-token context length as of July 2026. This can support long prompts, document analysis, and agent memory, but applications still need retrieval checks, prompt budgeting, and output validation.
How much does Hy3 Preview cost on Gate.AI?
As per the Gate.AI listing, Hy3 Preview costs \$0.18 per 1M input tokens, \$0.06 per 1M cache-read tokens, and \$0.60 per 1M output tokens as of July 2026. Cache-write pricing is not specified in the available listing.
How can developers access Hy3 Preview?
Developers can access Hy3 Preview through Gate.AI using the model ID tencent/hy3-preview with Gate.AI’s OpenAI-compatible API. Tencent also provides open-weight access through GitHub, Hugging Face, ModelScope, and GitCode, plus deployment guidance for vLLM and SGLang.
What is Hy3 Preview useful for?
Hy3 Preview is suitable for long-context reasoning, coding assistance, complex planning, document analysis, and multi-step agent workflows. It should be evaluated on private tasks before production use, especially where accuracy, latency, cost, or safety controls matter.
