Model library / tencent
Hy4 preview is a 780 billion-parameter mixture-of-experts language model released by Tencent for productivity-focused tasks.
8 of 256 experts active per token
Context length in tokens, as published
Weights at the released precision
Licence declared on the repository
Overview
Hy4 preview is a new-generation mixture-of-experts (MoE) model from Tencent's Hy team. It comprises 780 billion total parameters, with 256 experts per layer and eight active experts per token. The model is positioned for software engineering, office automation, game development and scientific research, aiming to improve long-horizon reasoning and document generation.
In practice the model requires substantial resources: it holds 780 billion parameters in bfloat16, activates roughly 49 billion per token, and supports a context window of 1,048,576 tokens. It is distributed under the Apache-2.0 licence, making it open-source and permissively licensed for commercial use. Deployment typically involves MoE-aware serving stacks such as vLLM or SGLang.
Specification
| Repository | tencent/Hy4-preview |
| Publisher | tencent |
| Published | 27 August 2026 |
| Architecture | HYV4ForCausalLM |
| Model type | hy_v4 |
| Parameters | 779,960,992,733 |
| Experts | 256 total, 8 active per token |
| Layers | 78 |
| Hidden size | 6,144 |
| Attention heads | 64, 8 key/value heads |
| Context length | 1,048,576 tokens |
| Vocabulary | 120,832 tokens |
| Weight format | bfloat16 |
| Licence | apache-2.0 |
Memory
Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.
| Precision | Weights | Notes |
|---|---|---|
| bf16 | 1452.8 GiB | Full precision as released |
| fp8 | 726.4 GiB | 8-bit, near-lossless on most models |
| int4 | 363.2 GiB | 4-bit, smallest footprint |
The cache costs 156 KiB per token at 16-bit, so the full 1,048,576-token context of one request needs about 156.0 GiB. Concurrency multiplies that number, not the weights.
That figure is an upper bound: this model uses compressed key-value attention, so the real cache is a fraction of the figure above.
Hardware
At the released precision, with 90% of the card given to the server and 2.5 GiB kept for the runtime. The context column is what is left for the cache on that setup.
| GPU | Cards needed | Free for cache | Context it holds |
|---|---|---|---|
| RTX 4090 24GB | more than 8 | — | — |
| RTX 5090 32GB | more than 8 | — | — |
| L40S 48GB | more than 8 | — | — |
| A100 80GB | more than 8 | — | — |
| H100 80GB | more than 8 | — | — |
| RTX PRO 6000 96GB | more than 8 | — | — |
| H200 141GB | more than 8 | — | — |
Serving
A vLLM launch line for the shape above. Check the model card for a runtime the publisher recommends.
vllm serve tencent/Hy4-preview \ --tensor-parallel-size 8 \ --max-model-len 1048576 \ --gpu-memory-utilization 0.90
At the released precision the weights do not fit on eight of the largest cards above, so this needs more than one node, or a quantised build.
Questions
The weights take about 1453 GiB at the released precision. No single card in the table above holds that, so it needs several GPUs or a lower precision.
The repository declares apache-2.0. Read the licence text before commercial use: the name alone does not tell you what is allowed.
The configuration allows 1,048,576 tokens. Whether the whole window is usable depends on the memory left for the cache, which the hardware table works out per card.
This page is a reference, and Aiqre does not serve this model today. We deploy open-weight models on dedicated EU hardware on request, so if you want an OpenAI-compatible endpoint for it, email [email protected] with the model name and roughly what volume you expect.
Also in the library
Nex-N2.5-Pro is a 396.8 billion-parameter MoE language model designed for long-horizon, agentic tasks with multimodal and compute…
GigaChat 3.5 Reasoning is a 438 billion-parameter mixture-of-experts language model trained with online RL for improved reasoning…
GLM-5.3 is a 753.3 billion-parameter open-weight mixture-of-experts language model for coding and security tasks.
Hosting
Tell us the model and roughly what volume you expect. We reply with a price and an OpenAI-compatible endpoint, with your prompts never stored.