Join our Newsletter — 33% off our NHI Course
Home› Guides› Hunyuan Hy3: Specs, Context Window, Pricing, Benchmarks and…
Guide AI Platforms

Hunyuan Hy3: Specs, Context Window, Pricing, Benchmarks and Availability

PDF for newsletter subscribers · free to read online

← All guides
By NHI Mgmt Group, under the editorial direction of Lalit Choda Updated 18 November 2026 15 min read
On this page

Hunyuan Hy3 is an open-weight Mixture-of-Experts large language model from Tencent, released on 6 July 2026, with 295 billion total and 21 billion active parameters and a 256K-token context window. This profile collects the published facts about it in one place: model ID, architecture, context and output limits, reasoning modes, prices in yuan for online and batch calls, where it runs, benchmark scores Tencent reports and how it relates to the rest of the Tencent Hy family. It is written for developers, architects, buyers and governance teams who need checked numbers rather than launch coverage. Every fact comes from Tencent's newsroom, Tencent Cloud TokenHub documentation or Tencent's Hugging Face model cards, listed under Standards and references. Where Tencent publishes nothing on a point, the page says so.

Last checked: 2026-11-17.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Tencent's public Hy3 model card and Tencent Cloud TokenHub documentation as of November 2026. Model details change, so confirm against Tencent's current documentation.

Key takeaways

  • Released 6 July 2026 under Apache 2.0. Tencent's announcement dates the launch to 6 July 2026, after a preview on 23 April, and the final weights carry the Apache License 2.0; the preview used the Tencent Hy Community License Agreement.
  • 295B total, 21B active, 256K context. The Hugging Face card lists 192 experts with 8 active per token, 80 layers plus one multi-token prediction layer, and a 256K context length.
  • 192K in, 128K out on TokenHub. Tencent Cloud's model list gives hy3 a 256k context window, 192k maximum input and 128k maximum output, with deep thinking, structured output, function calling and caching.
  • 1 yuan in, 4 yuan out per million tokens. TokenHub online pricing is 1 yuan input, 4 yuan output and 0.25 yuan for cache hits, with batch calls at half those rates.
  • No offline date. Tencent Cloud's model list shows no offline date for hy3 as of 16 November 2026, while hy4-preview is listed to go offline on 7 December 2026.

What is Hunyuan Hy3?

Hy3 is the third-generation general language model in Tencent's foundation model programme. Tencent's July 2026 announcement says the family formerly called Tencent Hunyuan is now known as Tencent Hy, so "Hunyuan Hy3" and "Tencent Hy3" refer to the same model. The announcement describes Hy3 as a "hybrid fast-and-slow-thinking model built on a Mixture-of-Experts (MoE) architecture", meaning one checkpoint can answer directly or produce a reasoning trace first, depending on the mode the caller sets.

Tencent states that it began rebuilding its training infrastructure in late January 2026 and that Hy3 preview was the first model trained on that rebuilt stack. The announcement names software development, office productivity, financial modelling, front-end design and game production as the areas of strongest improvement, and says Hy3 matches flagship-model intelligence from models with two to five times its parameter count; it gives no benchmark figures for that claim. Inside Tencent, Hy3 powers WorkBuddy, CodeBuddy, the Yuanbao app's Agent feature, Marvis Agent, ima, Weixin Official Accounts customer-service assistants and an in-game assistant for Path of Exile: Advent on WeGame. For the full Tencent line-up, see Tencent Hunyuan Models.

Key facts at a glance

As of 2026-11-17. Sources: Tencent Hy3 Hugging Face card, Tencent newsroom announcement, Tencent Cloud TokenHub model list and pricing pages.

FieldValueSource
ProviderTencent (Tencent Hy Team)Hugging Face card; Tencent newsroom
API IDhy3 (TokenHub); served model name hy3 in the card's vLLM and SGLang examplesTokenHub model list; Hugging Face card
Release date2026-07-06 (preview 2026-04-23)Tencent newsroom; Hy3-preview card
StatusCurrent; no offline date listedTokenHub model list
Context window256K tokensHugging Face card; TokenHub model list
Max input (TokenHub)192K tokensTokenHub model list
Max output (TokenHub)128K tokensTokenHub model list
Knowledge cutoffNot stated by Tencent on the pages checked (17 Nov 2026)Hugging Face card; TokenHub model list
Input modalitiesTextHugging Face card
Output modalitiesTextHugging Face card
Reasoning modesno_think (default), low, high, set with reasoning_effortHugging Face card
Open weightsYes: Hy3 (BF16) and Hy3-FP8Hugging Face card
LicenceApache License 2.0Hugging Face card; Tencent newsroom
Parameters295B total, 21B active, plus 3.8B in the MTP layerHugging Face card
Fine-tuningFinetuning guide and GRPO reinforcement learning recipe in the repositoryHugging Face card

The Hugging Face page metadata shows 299B parameters, while the card text and architecture table both state 295B; this page uses the card figure.

Release and version history

As of 2026-11-17. Dates from Tencent's newsroom, the Hy3 and Hy3-preview Hugging Face cards and the TokenHub model list.

DateEventSource
Late January 2026Tencent begins rebuilding its model training infrastructureTencent newsroom
2026-04-23Hy3 preview weights open-sourced on Hugging Face, ModelScope and GitCode under the Tencent Hy Community License Agreement, with a Hy3 preview-Base checkpointHy3-preview card; Tencent newsroom
2026-07-06Hy3 released: API on TokenHub, Hy3 and Hy3-FP8 weights on Hugging Face and ModelScope (also GitCode and CNB) under Apache 2.0Tencent newsroom; Hy3 card
2026-11-16TokenHub model list and pricing pages last updated; hy3 listed with no offline dateTokenHub model list and pricing

Tencent states that average daily token consumption of the model rose twentyfold between the preview launch and the July release, and that the number of WorkBuddy users choosing Hy3 preview grew sixfold. Hy3 preview is not listed as a model on the current TokenHub model list or pricing page; the API serves hy3.

How much does Hunyuan Hy3 cost?

As of 2026-11-17. Chinese yuan (CNY) per million tokens. Source: TokenHub model pricing page (updated 2026-11-16) and TokenHub product page, which states it shows Guangzhou-region pricing.

ItemPriceNotesSource
Online (standard)Input: 1 yuan; Output: 4 yuan; Cache hit: 0.25 yuanCache write not offered as a separate itemTokenHub pricing page
BatchInput: 0.5 yuan; Output: 2 yuan; Cache hit: 0.125 yuanHalf the online ratesTokenHub pricing page
Free quotaHy3 not listed in the free-quota tableThe table covers other models such as Hy4 and Hy4 flashTokenHub product page
DiscountsNo Hy3 discount listedTime-limited 50% discounts apply to Hy4 and Hy4 flashTokenHub pricing page
Self-hostedNo per-token fee; weights free under Apache 2.0Infrastructure cost is the customer'sHugging Face card
  • Billing model: TokenHub bills pay-as-you-go per million tokens, with input, output and cached input as separate line items; settlement is hourly with an expected two to three hour delay.
  • Interrupted requests: the billing page states that tokens generated before an interruption are billed, and tasks that return no usable result are not billed.
  • Prepaid plans: TokenHub also offers monthly token plans, including a Coding Plan, a General Token Plan, a Hy Token Plan and two enterprise options; their prices are on separate pages not checked for this profile.
  • Third-party hosts: the Hy3 card lists DeepInfra as an inference provider and HuggingChat; their prices are set by those services and are not covered here.

Worked example at online rates: a request with 100,000 input tokens and 10,000 output tokens costs 0.10 yuan for input plus 0.04 yuan for output, 0.14 yuan in total. In high reasoning mode the reasoning trace is generated output, so plan budgets on output tokens including reasoning. See LLM Pricing Tracker for other providers' current rates.

Context window and output limits

The Hugging Face card gives a 256K context length for the model. TokenHub's model list splits that window into a maximum of 192k input tokens and 128k output tokens, so a request cannot use the full 256k for the prompt. Self-hosted deployments are limited by the serving engine settings and GPU memory rather than by TokenHub's split.

LimitValueNotesSource
Context window256K tokensModel capabilityHugging Face card; TokenHub model list
Max input (TokenHub)192K tokensPer requestTokenHub model list
Max output (TokenHub)128K tokensPer requestTokenHub model list
Rate limits (TokenHub)Not stated by Tencent on the pages checked (17 Nov 2026)NoneTokenHub model list; API usage page
Vocabulary120,832 tokensTokeniser sizeHugging Face card
Recommended self-hosting hardware8 GPUs; H20-3e or GPUs with larger memory suggestedTensor parallel size 8 in examplesHugging Face card

For how context windows interact with retrieval and cost, see Context Windows Explained.

Modalities and tools

Hy3 is text in, text out. Tencent's vision understanding is offered through other models in the family, and the TokenHub page states that Hy4 is the first Hy flagship with vision understanding.

  • Reasoning modes: reasoning_effort set to no_think (the default, direct response), low, or high (deep chain-of-thought, which the card recommends for maths, coding and reasoning).
  • TokenHub capabilities: the model list names deep thinking with retained thinking, structured output, function calling and caching for hy3.
  • API protocols: TokenHub supports both the OpenAI and Anthropic API protocols, and the OpenAI SDK can be used directly; the card describes an OpenAI-compatible chat completions API.
  • Recommended sampling: temperature 0.9 and top_p 1.0.
  • Self-hosted serving: vLLM with the hy_v3 tool-call and reasoning parsers and two-token multi-token prediction speculative decoding; SGLang (transformers 5.6.0 or later) with the hunyuan parsers and EAGLE speculative decoding; Docker Model Runner; community quantisations for llama.cpp, Ollama and LM Studio.
  • Agent tooling: the card reports that SWE-bench Verified accuracy varies by less than 4% across the CodeBuddy, Cline and KiloCode scaffolds.

Function calling gives the model reach into connected systems, so every tool is a credentialed integration. Self-hosted endpoints exposed with OpenAI-compatible APIs need authentication in front of them; see the vLLM, Triton and NVIDIA NIM Inference Server Security Guide.

Where is Hunyuan Hy3 available?

As of 2026-11-17. Sources: Tencent newsroom, Hy3 Hugging Face card, TokenHub API usage page.

ChannelDetailsSource
Tencent Cloud TokenHub APIIdentifier: hy3; Guangzhou base URL tokenhub.tencentmaas.com (mainland China scheduling) and Singapore base URL tokenhub-intl.tencentmaas.com (global scheduling)TokenHub API usage page; model list
Open weightsHugging Face (tencent/Hy3, tencent/Hy3-FP8), ModelScope, GitCode, CNBHugging Face card; Tencent newsroom
Hosted inferenceDeepInfra (Hugging Face inference provider), HuggingChatHugging Face card
Third-party platformsAnnounced as coming progressively to OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode and Cherry StudioTencent newsroom
Tencent productsWorkBuddy, CodeBuddy, Yuanbao, Marvis Agent, ima, Weixin Official Accounts assistantsTencent newsroom
Official siteaistudio.tencent.comHugging Face card

TokenHub calls must target the region where the service is enabled; the API usage page states that cross-region and cross-site calls are not supported, and it lists backup base URLs on the .cn domain for both regions.

Access restrictions

Hy3 is not a restricted-access model. The final Hugging Face card states no territory restriction, and the weights are published under Apache 2.0, which permits commercial use. The earlier Hy3 preview used the Tencent Hy Community License Agreement, so teams that downloaded the April preview should check which licence applies to the checkpoint they hold. API access requires a Tencent Cloud account with TokenHub enabled in the chosen region and an API key sent in an Authorization: Bearer header; the API usage page states that a missing or invalid key returns HTTP 401. For licence terms across open models, see the Open-Weight Model Licence Tracker.

Benchmarks (provider-reported, dated)

As of 2026-11-17. All figures are provider-reported on Tencent's Hugging Face cards; the card text does not state evaluation settings (for example tools or reasoning mode) for each score, and items marked with an asterisk on the card are marked the same way here. Independent leaderboard results are not included because none were checked against their own published pages for this profile.

EvaluationHy3Comparator in sourceSource and date
GPQA Diamond90.4Hy3 preview: 87.2Hy3 card; Hy3-preview card
SWE-bench Verified (resolved)78Hy3 preview: 74.4Hy3 card; Hy3-preview card
SWE-bench Multilingual (resolved)75.8None in textHy3 card
Terminal-Bench 2.171.7*Hy3 preview on Terminal-Bench 2.0: 54.4 (different version)Hy3 card; Hy3-preview card
Humanity's Last Exam (HLE)53.2Hy3 preview: 30Hy3 card; Hy3-preview card
Apex Agents25.6None in textHy3 card
Deep SWE28None in textHy3 card
Blind expert evaluation (270 experts, score out of 4)2.67GLM-5.1: 2.51Hy3 card

The Hy3 card also reports internal quality measures: hallucination rate falling from 12.5% to 5.4%, commonsense error rate from 25.4% to 12.7% and multi-turn issue rate from 17.4% to 7.9%. For the base model, the Hy3 preview card reports MMLU (5-shot) 87.42, MATH (4-shot) 76.28, GSM8K (4-shot) 95.37 and LiveCodeBench-v6 (1-shot) 34.86 for Hy3 preview-Base. Further comparison charts on both cards are images and are not transcribed here. See How to Read LLM Benchmarks.

Model card and system card

Tencent publishes a Hugging Face model card for Hy3 and a separate card for Hy3 preview; no separate system card or safety report is linked from the pages checked. What the Hy3 card states:

  • Architecture: 80 layers plus one MTP layer, 64 attention heads with grouped-query attention and 8 key-value heads, head dimension 128, hidden size 4,096, intermediate size 13,312, 192 experts with top-8 routing, BF16 precision.
  • Training and tuning: a finetuning guide and a GRPO reinforcement learning recipe using verl, with Megatron-LM (via NVIDIA Megatron-Bridge conversion) and vLLM rollout; Tencent's AngelSlim toolkit for compression.
  • Quality work: the card describes anti-hallucination and fact-conflation work with the internal measures listed in the benchmarks section; it has no dedicated safety, red-teaming or misuse section.
  • Not stated: knowledge cutoff, training data size and composition, and supported languages (the README is published in Chinese and English).

Because the card does not publish safety evaluations, teams deploying Hy3 for external users should run their own red-team and content-safety testing. See Model Cards and System Cards Explained.

Data handling

Data handling depends on how Hy3 is used:

  • Self-hosted weights: prompts and outputs stay in the customer's own environment; retention, logging and access control are the operator's responsibility.
  • TokenHub API: retention periods and training use of API data are not stated by Tencent on the TokenHub pages checked (17 Nov 2026). Confirm terms in the Tencent Cloud service agreement before sending personal or regulated data.
  • Region: the Guangzhou endpoint uses mainland China scheduling and the Singapore endpoint global scheduling; region choice affects where requests are processed and which data transfer rules apply. See the China Cross-Border Data Transfers.
  • Third-party hosts: DeepInfra, HuggingChat and other platforms apply their own data terms.

TokenHub API keys are non-human identities with direct spend and data exposure. Store them in a secrets manager and scope them per project; see the LLMjacking Guide and the Tencent Cloud Security Guide.

Deprecation status

As of 16 November 2026, Tencent Cloud's TokenHub model list shows no offline date for hy3. The same list does publish offline dates for other models, which shows the format to watch for.

Model or itemDetailsSource
hy3No offline date listed (2026-11-16)TokenHub model list
hy4-previewOffline 2026-12-07TokenHub model list
Hy3 previewNot listed on the TokenHub model list or pricing page; weights remain on Hugging FaceTokenHub model list; Hy3-preview card

Open weights do not expire, so a self-hosted Hy3 deployment is unaffected by API retirement. See the AI Model Retirement Guide.

How Hunyuan Hy3 differs from other models in its family

As of 2026-11-17. Source: TokenHub pricing page and model list (updated 2026-11-16), Hy3 and Hy3-preview cards. Models listed by generation. Prices are online CNY per million tokens.

ModelAPI IDInput / output / cache hitDiscount statedOffline date
Hy3 previewNot listed on TokenHubNot listed on TokenHubNoneNot listed
Hy3hy31 / 4 / 0.25NoneNone listed
Hy4 flashhy4-flash0.5 / 2 / 0.0250% until 2026-11-30 12:00 Beijing timeNone listed
Hy4 previewhy4-preview6 / 18 / 0.3None2026-12-07
Hy4hy46 / 18 / 0.350% until 2026-12-31 23:59:59 Beijing timeNone listed

Hy3 and Hy3 preview share the same 295B total, 21B active architecture and 256K context; the documented differences are the licence (Apache 2.0 for Hy3, the Tencent Hy Community License Agreement for the preview) and the benchmark scores above. Hy4 is the newer flagship generation at six times Hy3's online input price, and Hy4 flash is the lower-cost Hy4 option. Architecture details for the Hy4 models are on the Tencent Hunyuan Models hub.

Practitioner checklist

  • Pin the model ID hy3 in code, and for self-hosting record the exact Hugging Face revision of tencent/Hy3 or tencent/Hy3-FP8 you deployed.
  • Confirm which licence applies to the checkpoint you hold: Apache 2.0 for Hy3, the Tencent Hy Community License Agreement for the April preview.
  • Choose the TokenHub region (Guangzhou or Singapore) deliberately, record the data transfer basis, and remember that cross-region calls are not supported.
  • Plan prompts within the 192k input limit on TokenHub, not the 256K model context.
  • Set reasoning_effort per workload: no_think for short tasks, high only where measured accuracy gains justify the extra output tokens.
  • Use batch pricing for non-interactive jobs, since it halves the online rate.
  • Run your own safety and prompt-injection testing before external deployment, because the card publishes no safety evaluations.
  • Store TokenHub API keys in a secrets manager, separate keys by environment, and put authentication in front of any self-hosted OpenAI-compatible endpoint.
  • Check the TokenHub model list monthly for an offline date for hy3.

Standards and references

Related NHI Mgmt Group resources: Tencent Hunyuan Models, Open-Weight Frontier Models Compared, Open-Weight Model Security Guide, Self-Hosted LLM Security Guide, NVIDIA Nemotron 3.

PDF for subscribers

Take this guide with you

Every section, table and checklist in one PDF, with links to the sources and related guides. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.

Free weekly newsletter

Subscribe to the NHI & AI Identity Journal

The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.

Bonus 33% off our NHI Course when you subscribe.

PDF Downloads for Newsletter Subscribers

Get the guide

Enter your details and we’ll email you a download link.

The full guide is free to read online. The PDF is a thank you for subscribing.

Fill in all fields marked * and tick both boxes to continue.