Tencent Hunyuan Models: Every Model, Spec, Price and Release Date

Hunyuan, now branded Tencent Hy, is the family of large language models from Tencent, a Chinese technology company that serves them through the Tencent Cloud TokenHub API and its Yuanbao app and publishes many of them as open weights on Hugging Face; its general release Hy3 appeared on 6 July 2026 and Hy4 is now priced on TokenHub. This page lists the Hunyuan and Hy language models Tencent documents on its official pages, with model IDs, parameter counts, context windows, licences, published prices in yuan, release dates and status. It is written for developers, buyers and architects who need to know which Tencent model is which. Models are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from Tencent's newsroom, Tencent Cloud documentation or Tencent's Hugging Face model cards, cited in the sources list. Image, video, 3D and speech models are out of scope apart from a short mention.

Last checked: 2026-11-15.

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 Cloud's public TokenHub documentation and Tencent's Hugging Face model cards as of November 2026. Model details change, so confirm against Tencent's current documentation.

Key takeaways

• Hy4 is the newest model on TokenHub. The TokenHub page describes the formal Hy4 release as strengthening agent and long-horizon task abilities and adding vision understanding for the first time, priced at 6 yuan input and 18 yuan output per million tokens.
• Hy3 is the current mid-price open model. Released on 6 July 2026 after an April preview, Hy3 is a 295B total, 21B active Mixture-of-Experts model with a 256K context under Apache 2.0, priced at 1 yuan input and 4 yuan output per million tokens.
• Hy4 preview weights are open. Tencent's Hugging Face card lists Hy4 preview at 770B total and 49B active parameters with a 1M-token context, under Apache 2.0.
• Hy4 is half price until the end of 2026. TokenHub states that Hy4 calls are settled at 50% of list price through 31 December 2026, 23:59:59 Beijing time.
• The older Hunyuan platform is closed to new models. The legacy Hunyuan billing page says features are moving to TokenHub and no new model capabilities will be added there; Hunyuan-A13B, T1 Vision and Turbo S Vision remain priced on it.

What is Tencent Hunyuan?

Hunyuan is Tencent's foundation model programme. Tencent's July 2026 Hy3 announcement says the family formerly called Tencent Hunyuan is now known as Tencent Hy, and newer models use Hy names (Hy3, Hy4) while older API IDs keep the hunyuan- prefix. Tencent serves its models through Tencent Cloud TokenHub, which its product page describes as a unified large-model service entry combining Tencent's in-house Hunyuan models with third-party models, and through consumer and work products including Yuanbao, WorkBuddy, CodeBuddy and ima.

Tencent publishes weights for many of its language models. Hy3, Hy4 preview and Hunyuan-A13B all have Hugging Face model cards, and the Hy3 and Hy4 preview weights are also mirrored on ModelScope, GitCode and CNB. Tencent's Hy3 announcement names software development, office productivity, financial modelling, front-end design and game production as target uses. Tencent also offers translation (Hy-MT2), role-play (Hy-Role), vision understanding (HY-Vision) and image, video, 3D and speech models; this page covers the general language models in depth and lists the others in the pricing section. Facts are as of 15 November 2026.

Current Hunyuan models at a glance

As of 2026-11-15. Source: Tencent Cloud TokenHub pricing and product pages, Tencent newsroom, Tencent Hugging Face model cards. Models are listed by release date, newest first; Hy4's release date is not stated on the official pages checked and it is placed first as the newest TokenHub entry. Category is the Tencent model line.

Hy3 and Hy4 preview cards describe text generation only; the TokenHub page states that Hy4 introduces vision understanding. Hy3 offers three reasoning modes (no_think, low and high), and Tencent's announcement describes it as a hybrid fast-and-slow thinking model. See Open-Weight vs Closed Models Explained.

Key facts for the main Hunyuan models

As of 2026-11-15. Identical fields for every model; "Not stated by Tencent" means not stated by Tencent on the pages checked (2026-11-15).

The Hy3 card adds 3.8B parameters in a multi-token prediction layer, and the Hy4 preview header shows 780B, which the card does not reconcile with the 770B figure. Hy3 has 192 experts with 8 active and 80 layers; Hy4 preview has 256 routed experts plus 1 shared expert with 8 active, 78 layers and Gated DSA attention.

Hunyuan model families and naming

• Hy generation number: Hy3 and Hy4 are successive flagship generations under the Tencent Hy brand. "Preview" marks an early checkpoint that is later replaced by the formal release.
• Legacy Hunyuan names: older API IDs use the Hunyuan prefix and a role, such as Hunyuan-a13b, Hunyuan-t1-vision (reasoning with vision) and Hunyuan-turbos-vision (fast model with vision).
• Parameter names: open checkpoints name active parameters, as in A13B (13B active of 80B total).
• Precision variants: Hy3-FP8 and Hy4 preview-FP8 are FP8-quantised checkpoints of the same instruct models; Hunyuan-A13B also ships FP8 and GPTQ-Int4 variants.
• Task models: Hy-MT2 (translation), Hy-Role (role-play), HY-Vision (multimodal understanding), plus Hy-Image, Hy-Video, Hy-World, HY-3D and Hy-ASR for generation and speech.

Architecture of the open checkpoints, as stated on each Hugging Face card (checked 2026-11-15):

The Hy3 card recommends sampling at temperature 0.9 and top_p 1.0, uses the hy_v3 tool-call and reasoning parsers in vLLM with two-token multi-token prediction speculative decoding, and the hunyuan parsers with EAGLE speculative decoding in SGLang. The Hy4 preview card lists official vLLM and SGLang Docker images and Docker Model Runner, with community quantisations for llama.cpp, Ollama and LM Studio.

When was each Hunyuan model released?

As of 2026-11-15. Dates are from Tencent's newsroom, Tencent Cloud pages and Tencent Hugging Face cards.

The Hy4 preview card gives no release date. For dated release history across providers see the LLM Release Tracker.

How much do Hunyuan models cost?

As of 2026-11-15. Source: Tencent Cloud TokenHub pricing (last updated 2026-11-09) and legacy Hunyuan billing page (last updated 2026-06-26). Yuan (CNY) per million tokens, pay-as-you-go; the TokenHub pricing page lists Guangzhou and Singapore. No context-length tiers apply to the Hy rows.

Batch figures are input, cached input and output. On TokenHub, HY-Vision-2.0-Instruct is 7.5 yuan input and 17.5 yuan output, and HY-Vision-1.5-Thinking and HY-Vision-Video are 3 and 9. The legacy platform gives a shared 1M-token free resource pack valid for one year after activation; TokenHub lists no free quota for Hy rows. Non-text Hy models on TokenHub are billed differently: Hy-Image-3.5-preview converts a 1K image to 15,000 tokens, or 0.15 yuan per image; HY-World-2.1-scene converts one scene to 8,000,000 tokens, or 80 yuan; and Hy-ASR-3.0-Preview speech recognition is 10 yuan per million tokens at 22 tokens per second of audio. TokenHub also offers guaranteed-resource and dedicated-deployment billing alongside pay-as-you-go, without published unit prices for Hy models on the pages checked. Compare providers in the LLM Pricing Tracker.

Where are Hunyuan models available?

As of 2026-11-15.

Tencent's Hy3 announcement reports that average daily token consumption rose twentyfold between the preview and the formal release, and that Yuanbao added an Agent function, free for everyday workplace needs, that generates office files. The TokenHub pricing page lists Guangzhou and Singapore regions, with the same Hy4, Hy3 and Hy-MT2-Plus prices in both.

Restricted-access Hunyuan models

What is the context window of Hunyuan models?

Hy4 preview has a 1M-token context according to its Hugging Face card. Hy3 has 256K, stated in both Tencent's announcement and its card. Hunyuan-A13B supports a native 256K (262,144 tokens), but its default config.json limits context to 32K, and self-hosters must set max_position_embeddings to 262144 to enable the full window. Hy4's context window on TokenHub is not stated on the pages checked, and no Tencent page checked states a maximum output length for any of these models.

Is any Hunyuan model open source?

Yes. Hy3, Hy3-FP8, Hy4 preview and Hy4 preview-FP8 are published under the Apache License 2.0, and Tencent's Hy3 announcement describes the licence as commercially friendly with open-source download and use for developers worldwide. Hunyuan-A13B (Pretrain, Instruct, FP8 and GPTQ-Int4) is published under the tencent-hunyuan-a13b licence, which is a Tencent licence rather than Apache 2.0; read its LICENSE file before commercial use. The Hy3 card recommends 8 GPUs such as H20-3e for serving and gives vLLM and SGLang settings. Self-hosting moves security, patching and access control to the operator; see the Open-Weight Model Security Guide and the Open-Weight Model Licence Tracker.

Model cards and system cards

The Hy4 preview card lists known issues: overly long reasoning on complex tasks and a tendency to over-verify its own work. See Model Cards and System Cards Explained.

Benchmarks

Provider-reported scores from Tencent's Hugging Face cards (checked 2026-11-15). Tencent chose the benchmarks and settings; scores are not independently verified.

The Hy3 card also reports SWE-bench Verified 78 and SWE-bench Pro 57.9, and a blind evaluation by 270 experts scoring Hy3 2.67 of 4. The Hy4 preview card reports Toolathlon 74.1 and an internal blind evaluation of 203 engineering tasks averaging 2.99. Hunyuan-A13B-Instruct reports AIME 2025 76.8, GPQA-Diamond 71.2, LiveCodeBench 63.9, IF-Eval 84.7, BFCL v3 78.3 and τ-Bench 54.7, and the Hunyuan-A13B base model reports MMLU 88.17 and MMLU-Pro 67.23. Scores on different cards use different settings and dates, so compare like with like. Cross-provider context is in Frontier AI Models Compared.

Deprecations and retirements

As of 2026-11-15.

A dated model retirement schedule or a minimum notice period is not stated by Tencent on the pages checked. Open-weight checkpoints remain downloadable independently of API status.

Data handling and security

• Training use: the TokenHub FAQ states that customer data is not used for scenarios beyond providing the service and is not used for training.
• Retention: a retention period or zero-retention option is not stated by Tencent on the pages checked.
• Compliance filings: the TokenHub page lists a deep synthesis algorithm filing with the Cyberspace Administration of China, a CAICT trusted AI rating and a software copyright registration for Hunyuan.
• Self-hosting: open weights run entirely on the operator's infrastructure; the operator owns access control, logging and patching.
• Jurisdiction: TokenHub is a Chinese cloud service; see the China Generative AI Measures and AI Labelling Rules Guide and China Cross-Border Data Transfers.
• Credentials: TokenHub API keys are non-human identities with spend and data exposure; see the LLMjacking Guide and LLM Data Retention and Training-Use Policies by Provider.

How to choose a Hunyuan model: factual criteria

Evaluate shortlisted models on your own data; provider benchmark tables use Tencent's chosen settings.

Practitioner checklist

• Use the exact TokenHub model ID (hy4, hy4-preview or hy3) and record the date adopted.
• Budget in yuan per million tokens, and note that the Hy4 50% settlement discount ends on 31 December 2026.
• Plan migration off the legacy Hunyuan platform; Tencent states no new model capabilities will be added there.
• Treat Hy4 preview as an early checkpoint with known issues; retest when moving to Hy4.
• For Hunyuan-A13B self-hosting, set max_position_embeddings if you need more than 32K context, and read the Tencent licence.
• Confirm Tencent Cloud data terms and cross-border data rules before sending personal data to TokenHub.
• Store TokenHub API keys in a secrets manager and set spend alerts.
• Run your own evaluations; provider-reported scores are not independent.

Standards and references

• Tencent Cloud: TokenHub billing (accessed 2026-11-15)
• Tencent Cloud: TokenHub product page (accessed 2026-11-15)
• Tencent: Hy3 release announcement (accessed 2026-11-15)
• Hugging Face: tencent/Hy4-preview model card (accessed 2026-11-15)
• Hugging Face: tencent/Hy3 model card (accessed 2026-11-15)
• Hugging Face: tencent/Hunyuan-A13B-Instruct model card (accessed 2026-11-15)
• Tencent Cloud: Hunyuan billing (legacy platform) (accessed 2026-11-15)

Related NHI Mgmt Group resources: ByteDance Doubao Seed Models, Alibaba Qwen Models, Baidu ERNIE Models, Open-Weight Frontier Models Compared.