Tencent Hunyuan and Hy models, generation by generation

Tencent's open LLM line from Hunyuan-A13B and 7B to Hy3 and Hy4 preview: sizes, MoE design, context, licenses, and which generation fits which job.

What the line is

Hunyuan is Tencent's family of open-weight language models. The first open releases carried the Hunyuan name. Later flagships are published by the Tencent Hy Team under the shorter Hy name (Hy3, Hy4 preview), and the Hugging Face tags for those models still include hunyuan. All of them are text models loaded through transformers, with vLLM and SGLang deployment instructions on the cards for Hy3 and Hy4 preview.

The line moved from a mid-size mixture-of-experts model and a dense 7B in mid-2025, to a 295B MoE in 2026, to a 770B MoE with a 1M-token window. License terms changed along the way, so check the license before you pick a generation.

Generations in order

Hunyuan instruct (June to July 2025)

Hunyuan instruct holds the two Hunyuan-branded instruct models covered here.

  • Hunyuan-A13B-Instruct was released on June 27, 2025. It is a fine-grained mixture-of-experts model with 80B total and 13B active parameters, grouped-query attention, and 256K tokens (262,144) of native context, with the default configuration set to 32K. The card lists the license as tencent-hunyuan-a13b, a Tencent-specific license rather than Apache 2.0 or MIT. It supports fast and slow thinking modes.
  • Hunyuan-7B-Instruct was released on July 30, 2025. It is dense, uses grouped-query attention, and natively supports a 256K window. Its card says it sits in a dense series alongside 0.5B, 1.8B and 4B variants and that it borrows training strategies from A13B. It also has fast and slow thinking modes, ships in BF16, and the card says FP8 and INT4 quantizations are supported. The license is the Tencent Hunyuan license in the model's GitHub repository, not a standard open-source license.

Hy3 (preview April 2026, full release July 2026)

Hy3 is a 295B-parameter mixture-of-experts model with 21B active parameters and a 3.8B multi-token-prediction layer. The card lists 80 layers, 64 attention heads (grouped-query, 8 KV heads), 192 experts with top-8 routing, a 256K context window, and BF16 weights.

Hy3 preview was announced on its card on April 23, 2026, and was described there as the first model Tencent trained on its rebuilt infrastructure. Its license field is other, with the badge naming the Tencent Hy Community license. The full Hy3 card says it followed the preview after feedback from more than 50 products and more post-training. Its Hugging Face repository was created on July 2, 2026, and its license is Apache 2.0. Same size, same architecture, same context as the preview, with a more permissive license. An FP8 variant is published too.

Hy4 (preview, August 2026)

Hy4 is the newest generation listed here. Hy4 preview is a 770B-parameter mixture-of-experts model with 49B active parameters per token. Its card describes 78 layers, 256 routed experts plus one shared expert with the top 8 routed experts active per token, and a 1M-token context window. For attention it uses Gated DeepSeek Sparse Attention, and it includes a native multi-token-prediction layer for speculative decoding. The license is Apache 2.0. The Hugging Face repository was created on August 27, 2026, and an FP8 variant is also published.

Treat the name literally: it is a preview, and the card has a Known Limitations section worth reading before you commit to it.

Which generation to use today

These recommendations come from sizes, context and license only. Check vendor-reported benchmarks on the cards yourself.

  • Smallest footprint with a long window: Hunyuan-7B-Instruct is the only dense model in the line here, with a 256K window. Read its license first, since it is Tencent's own.
  • Mid-size MoE you can self-host: Hunyuan-A13B-Instruct has 13B active parameters out of 80B total. Its weights still all have to be loaded, so the total, not the active count, drives memory.
  • Permissive license, large MoE: Hy3 is Apache 2.0 with 295B total and 21B active parameters. It is the Apache-licensed generation with the smaller total size, and it has a 256K window.
  • Longest context and largest model: Hy4 preview, with 1M tokens and 770B total parameters. It is Apache 2.0, but it is a preview and a multi-GPU deployment.
  • Commercial use of the preview releases: Hy3 preview uses the Tencent Hy Community license, while the final Hy3 is Apache 2.0, so prefer the final Hy3 unless you need the preview.

Side line: translation

Hy-MT2 is a family of multilingual translation models released on May 21, 2026, in three sizes: 1.8B, 7B and 30B-A3B (a mixture-of-experts model). The card describes them as fast-thinking models for complex real-world translation scenarios, covering translation among 33 languages, and lists the license as Apache 2.0. It is a separate line from the chat models above, so use it for translation work rather than as a general assistant.

You can run these models on Aquanode GPUs.

Hunyuan generations

Every Hunyuan generation we track, oldest first. Each hub lists all of its models with VRAM at native, FP8 and INT4 precision.

HubModelsSizesSmallest native VRAMCheapest live fit for itEst. $/hr
Hunyuan Instruct27.5B to 80.4B16.8 GBRTX A5000$0.176/hr
Hy33298.8B to 298.8B668 GBAMD MI300X × 4$10.52/hr
Hy42780.0B to 803.5B1743 GBNo live fit–

Smallest native VRAM is the lowest requirement among the publisher's own checkpoints in the hub, at the precision they are published in; see the methodology.

Other Hunyuan lines

Side lines from Tencent: specialised or companion models outside the main Hunyuan generations.

HubModelsSizesSmallest native VRAMCheapest live fit for itEst. $/hr
Hy-MT232.0B to 30.1B4.6 GBRTX 3070$0.088/hr

Run and fine-tune Hunyuan

Per-generation guides: engine support, chat template and context flags, then fine-tune memory for LoRA and QLoRA.

Hunyuan compared with other models

Side-by-side VRAM, context length and license for models of a similar size.

Best models by task

Hunyuan models appear on these ranked task pages.

Sources

Facts in the text above were read from these pages and are the publisher's own statements, not benchmarks run by Aquanode. Last reviewed 2026-10-07.

Start ComfyUI on a rented GPU.

© 2026 Aquanode. All rights reserved.

All trademarks, logos and brand names are the property of their respective owners.