What it is
A Tencent Hy Team open-weight language model using a 295B-parameter MoE architecture with 21B active parameters per token and a 3.8B MTP layer.
Independent Tencent Hy3 Model Guide
Tencent Hy3 Model: official access, open weights, API pricing, and safe deployment
Tencent Hy3 Model is Tencent Hy Team's 295B-parameter Mixture-of-Experts model, with 21B active parameters, a 256K context window, and a clear focus on reasoning, coding, tool use, and agent workflows. This guide keeps the practical bits in one place: where to read the official material, how to call the model, what it costs on Tencent Cloud TokenHub, and where you should double-check before shipping.
Fact check: July 16, 2026. Model specs, open-weight links, TokenHub price rows, API notes, and cloud status links were checked against official Tencent and repository sources.
Independent note:
This site is not Tencent, Tencent Cloud, or an authorized support channel. Manage accounts, API keys, billing, model downloads, and incidents only through Tencent's official pages.
Official Links
Where to go first
Use these links for anything that affects your account, wallet, deployment, license, or compliance posture. This guide can help you read the landscape, but the source of truth is still Tencent's own material.
| User need | Official destination | What it is good for |
|---|---|---|
| Research overview | Introducing Tencent Hy3 Model | The official Tencent Hy page for the release story, positioning, and product context. |
| Hy product home | Tencent Hy | The broader Hy entry point, including product navigation and public Hy material. |
| Open-source repository | Tencent-Hunyuan/Hy3 | Model card, deployment notes, fine-tuning references, RL training notes, and Apache 2.0 license. |
| Model weights | Tencent Hy3 Model on Hugging Face | BF16 model files, model card, community metadata, and common library entry points. |
| Quantized weights | Tencent Hy3 Model-FP8 on Hugging Face | FP8 quantized weights for teams trying to reduce serving cost and memory pressure. |
| China-friendly mirror | Tencent Hy3 Model on ModelScope | An official model distribution path that may be easier to use from mainland China networks. |
| API quick start | Tencent Cloud TokenHub Quick Start | How to create an API key and call OpenAI-compatible endpoints through TokenHub. |
| Tencent Hy3 Model calling guide | Hy calling guide | Tencent Hy3 Model protocol notes, model ID guidance, and examples for Chat Completions, Responses, and Anthropic-style calls. |
| API pricing | TokenHub model pricing | Current per-million-token rows for Tencent Hy3 Model, Tencent Hy3 Model preview, and other TokenHub language models. |
| Service health | Tencent Cloud Health Dashboard | Regional cloud status, maintenance notices, and incident tracking. |
| Privacy and legal terms | Tencent Cloud Privacy Policy | Privacy terms for Tencent Cloud services; pair this with your account region and service agreement. |
At A Glance
Tencent Hy3 Model in plain English
How to access it
Read the official Hy page, download weights from GitHub-linked model hubs, run it with vLLM or SGLang, or call hy3 through Tencent Cloud TokenHub.
Free and paid paths
The model weights are published under Apache 2.0. Managed API usage on TokenHub is metered by input, output, and cache-hit tokens.
Current model
Tencent Hy3 Model is the current public release. Tencent Hy3 Model preview is still useful for migration notes, but new builds should start with the hy3 model ID when available.
Open resources
Official assets include GitHub code and docs, Hugging Face weights, ModelScope mirrors, FP8 weights, deployment recipes, and training references.
Must verify
Prices, supported regions, rate limits, account eligibility, privacy terms, and license obligations should be checked directly before production rollout.
What is Tencent Hy3 Model?
Tencent Hy3 Model is Tencent's latest Hy-generation language model for builders who care about long context, tool reliability, coding, and agent-style workflows. The headline number is big, 295B total parameters, but the more important engineering detail is the MoE design: only 21B parameters are active for a given token, which is how Tencent Hy3 Model tries to stay capable without behaving like a runaway infrastructure bill.
The model is not just a chat toy with a nice benchmark chart. Tencent's own material emphasizes post-training with product feedback, better tool-call stability, stronger format following, lower hallucination rates in internal evaluations, and better retention across long or multi-turn conversations. Treat those claims as encouraging, not magical. Tencent Hy3 Model still needs evaluation against your prompts, your retrieval stack, your tools, and your failure budget.
What people use it for
- Agent workflows that need tool calling, multi-step planning, and stateful task follow-through.
- Software engineering work such as code review, bug analysis, test generation, and repository-scale assistance.
- Long-context reading, summarization, and extraction across large documents or multi-file inputs.
- Structured business work, including financial modeling notes, office automation, reports, and data cleanup.
- Research and experimentation with open-weight serving, fine-tuning, quantization, and reasoning controls.
Choose the right usage path
Start here when you want the release narrative, Tencent's own positioning, and the safest links out to model resources.
Use this when you want managed inference with OpenAI-compatible tooling, API keys, billing, monitoring, and cloud support.
Use vLLM or SGLang when you need control over infrastructure, data flow, latency tradeoffs, or custom evaluation pipelines.
Useful for quick experiments, but check which model ID, provider, region, logging policy, and price are actually being used.
Models
Model and version notes
Tencent Hy3 Model has a small public family, but the names matter. Use the final Tencent Hy3 Model release for new work, keep Tencent Hy3 Model preview in mind when reading older examples, and consider Tencent Hy3 Model-FP8 if serving cost or GPU memory is your main constraint.
| Model or version | Positioning | Access | Verification note |
|---|---|---|---|
| Tencent Hy3 Model | Current 295B MoE instruct model for reasoning, coding, long context, and agent workflows. | GitHub-linked model hubs, local serving, and Tencent Cloud TokenHub model ID hy3. |
Checked July 16, 2026 against Tencent Hy, GitHub, Hugging Face, and TokenHub pricing docs. |
| Tencent Hy3 Model-FP8 | FP8 quantized Tencent Hy3 Model instruct model for teams trying to lower memory footprint and serving cost. | Hugging Face, ModelScope, GitCode, and CNB links from the official repository. | Evaluate quality and latency against your own prompts before replacing BF16 in production. |
| Tencent Hy3 Model preview | Earlier Tencent Hy3 Model-generation preview release, still visible in docs and pricing rows. | Older model hubs and TokenHub model ID hy3-preview where still supported. |
Treat as a migration reference unless a provider specifically tells you to use it. |
| Hy multimodal models | Separate Tencent Hy vision, image, video, and 3D models available through TokenHub. | TokenHub multimodal documentation, not the Tencent Hy3 Model text-model repo. | Do not assume Tencent Hy3 Model itself is the image, video, or 3D model; check the model ID and endpoint. |
API & Pricing
TokenHub API snapshot
TokenHub exposes language models through OpenAI-compatible APIs and also supports Anthropic-style Messages and OpenAI Responses patterns in its documentation. For Tencent Hy3 Model, the practical starting point is a Bearer API key, base URL https://tokenhub.tencentmaas.com/v1, and model ID hy3.
| API model | Input price | Output price | Cache hit | Notes |
|---|---|---|---|---|
hy3 |
1 RMB / 1M tokens | 4 RMB / 1M tokens | 0.25 RMB / 1M tokens | Guangzhou TokenHub language-model row checked July 16, 2026. Confirm region, account, and billing plan in the console. |
hy3-preview under 16K input |
1.2 RMB / 1M tokens | 4 RMB / 1M tokens | 0.4 RMB / 1M tokens | Preview-era pricing remains documented; use only when your provider still routes to the preview model. |
hy3-preview 16K to 32K input |
1.6 RMB / 1M tokens | 6.4 RMB / 1M tokens | 0.6 RMB / 1M tokens | Budget long-context runs separately; output cost climbs in the longer preview bands. |
hy3-preview 32K+ input |
2 RMB / 1M tokens | 8 RMB / 1M tokens | 0.8 RMB / 1M tokens | Use this row only as a documented preview comparison, not as a guarantee for final Tencent Hy3 Model. |
Privacy & Security
Use Tencent Hy3 Model like a professional, not like a pastebin
Do not paste secrets
Keep passwords, API keys, signing certificates, private customer data, financial records, health data, and unredacted legal documents out of prompts unless your policy explicitly permits it.
Verify high-impact answers
Use Tencent Hy3 Model as a draft partner for law, medicine, finance, safety, hiring, education, and security work, not as the final authority.
Control production access
Scope API keys, isolate environments, cap budgets, log request IDs, gate tool calls, and require review before agents touch money, data deletion, or customer-visible actions.
Private serving still needs ops
Local deployment reduces third-party data exposure, but it does not solve network security, access control, logs, storage, monitoring, patching, or internal misuse by itself.
Solutions
Developer and enterprise use cases
Tencent Hy3 Model is most interesting where a model has to keep a lot in its head, call tools without falling apart, and stay cost-aware. The best deployments usually combine Tencent Hy3 Model with retrieval, policy checks, observability, and boring but beautiful approval gates.
Application development
Use Tencent Hy3 Model for chat, extraction, classification, structured output, code review, test drafting, report generation, and workflows where tool-call stability matters.
RAG and knowledge bases
Pair the 256K context window with retrieval, citations, permission filters, tenant isolation, evaluation sets, freshness checks, and retention rules.
Agents and automation
For multi-step agents, define tool allowlists, approval points, rollback paths, rate limits, cost budgets, request tracing, and failure escalation.
Editorial Trust
How this guide checks facts
- Model facts come from Tencent Hy, Tencent's public release article, the Tencent-Hunyuan/Hy3 repository, Hugging Face, and ModelScope links.
- API facts come from Tencent Cloud TokenHub documentation, including quick start, Hy calling guide, model pricing, error formats, and monitoring pages.
- Time-sensitive details such as price, supported regions, service status, rate limits, subscriptions, and privacy terms are marked as items to re-check.
- Editorial copy is intentionally interpretive. It explains what the facts mean for builders without pretending to be Tencent support.
FAQ
Common questions
What is Tencent Hy3 Model?
Tencent Hy3 Model is Tencent Hy Team's 295B-parameter MoE language model with 21B active parameters and a 256K context window. It is positioned for reasoning, coding, long-context work, and agent workflows.
What is the official Tencent Hy3 Model website?
The official research page is https://hy.tencent.com/research/hy3. The broader Hy home is https://hy.tencent.com/, and the official open-source repository is https://github.com/Tencent-Hunyuan/Hy3.
Is Tencent Hy3 Model free?
The open weights are published under Apache 2.0, so downloading and self-hosting can be free apart from your own infrastructure cost. Managed TokenHub API calls are paid by token usage or covered by eligible token plans.
Does Tencent Hy3 Model support an API?
Yes. Tencent Cloud TokenHub documents Tencent Hy3 Model with model ID hy3 and OpenAI-compatible access. The base URL shown in Tencent's quick-start material is https://tokenhub.tencentmaas.com/v1.
Can Tencent Hy3 Model run locally?
Yes, but it is a large model. Tencent's repository gives vLLM and SGLang recipes and recommends serious multi-GPU serving hardware for full-size deployment.
Is Tencent Hy3 Model suitable for enterprise data?
Potentially, but only after a security review. Decide whether you are using TokenHub, private serving, or a third-party provider, then verify privacy terms, region, logging, retention, access controls, and incident handling.
Content Map
Future article map
These are the next pieces I would build after the homepage: short, useful, source-aware articles that answer the questions real builders search for before they put Tencent Hy3 Model into a product.