Tencent Hy4, explained: what Tencent shipped, what it costs, and what running it yourself would take

On 4 November 2026 Tencent's Hy team published "Introducing Hy4". The model is on Tencent Cloud, on OpenRouter as tencent/hy4, and in Tencent's own products: CodeBuddy (a coding assistant), WorkBuddy (an office assistant connected to WeCom, Tencent's work chat), Yuanbao (its chatbot) and ima (its knowledge-base app).

Hy4 is the finished version of Hy4 preview, which Tencent released on 27 August with its weights on Hugging Face under the Apache 2.0 licence, a licence that allows commercial use. Tencent says Hy4 "substantially scales up both pre-training and post-training". Pre-training is the first, long round of learning from text; post-training is the later rounds that teach the model to follow instructions and use tools.

Three things are new. Hy4 reads images (OpenRouter lists its input as text and image; the preview is text only). It can use a computer, both through the command line and through graphical programs, with targeted training on tools such as Blender and CAD software. And Tencent improved its search answers for finance and office work, which is where most of its own numbers come from.

The weights are promised, not delivered. Tencent writes that it plans to open-source them, has partnered with the vLLM and SGLang projects (free software for serving models), and will publish "a comprehensive system card" with safety results when it does. It gives no date. Tencent also says a smaller, cheaper Hy4 flash is coming "in the near future".

Tencent's Hy4 post does not give the model's size or design. The preview's model card does, and it is the best guide we have until the weights come out. Hy4 may differ: it adds an image reader at least, and Tencent has not said whether the rest changed.

The preview is a mixture of experts. Each of its 78 layers, except the first, holds 256 small sub-networks called experts, and for every token a router (a small part of the model that chooses) picks 8 of them, plus 1 shared expert that always runs. So about 49 billion of its 770 billion parameters work on each token. That keeps each token fast, but all 770 billion must sit in GPU memory, because any expert may be picked next.

Attention is the step where each token looks back at the tokens before it. The preview uses what Tencent calls Gated DeepSeek Sparse Attention. A small scorer, called the indexer, rates the earlier tokens and picks the 2,048 most useful ones, and only those are read in full. To make this work, each layer keeps a compressed record of every earlier token: 512 numbers, plus 64 for its position. That record is the KV cache (key-value cache), and it is what limits how many long conversations one server can hold.

The preview also has one extra layer that guesses the next few tokens ahead, so the server can check several at once and write faster. Tencent's own serving commands turn it on, with three guessed tokens per step.

Read next: Reading a model name: what 27B, FP8, A3B and Instruct mean

Open weights are why a company would pick this model over a closed one: you can run it on your own servers, keep your data there, and change it. Hy4's weights are not out and its size is not published, so these numbers use the open preview. If Hy4 keeps the preview's size they are a fair guide, but the image reader may add memory, and Tencent has not said how much.

The preview's 8-bit files are 814 GB. Tencent's own vLLM and SGLang commands use tensor parallelism across 8 GPUs: the model is cut into 8 slices, one per GPU, and all 8 work on every token together. That puts about 102 GB of weights on each GPU. An NVIDIA H100 has 80 GB, so eight H100s (80 GB each, 640 GB in all) cannot hold the preview. An H200 has 141 GB and a B300 has 288 GB, so both can.

Next, the KV cache from earlier. From the preview's config file, in 16-bit numbers, each token needs 96,384 bytes: 89,856 for the attention record across 78 layers, 5,376 for the indexer's keys on the 21 layers that run their own indexer (the other 57 reuse an earlier layer's choice of tokens and store nothing for it), and 1,152 for the extra guessing layer. That is about 101 GB for one conversation that fills the whole million tokens, or 3.2 GB for one of 32,768 tokens.

Here is the catch. With this kind of compressed attention and pure tensor parallelism, the cache is not split across the 8 GPUs: every GPU keeps its own full copy. So what matters is what is left on ONE GPU. On an H200 that is about 39 GB: no full-length conversation, and about 12 conversations of 32K at once. On a B300 it is about 186 GB: 1 full-length conversation, or about 58 of 32K. Serving setups that give each GPU its own requests instead (SGLang calls this data-parallel attention) keep each conversation's cache on one GPU only, so the server holds several times more 32K conversations. A full million-token conversation still has to fit on a single GPU. Storing the cache in 8-bit numbers roughly halves it. The serving software also needs working memory of its own.

Read next: Scaling and GPU infrastructure: serving models without burning money

The facts about Hy4 come from Tencent's post "Introducing Hy4" and OpenRouter's public model and endpoint lists, read on 5 November 2026. The design and hardware facts come from the Hy4 preview's model card, config file and licence on Hugging Face. All are linked below.

I did not send a prompt to Hy4: this project has no key for OpenRouter or Tencent Cloud, and the weights are not out. So I cannot tell you how good its answers are, how long its reasoning runs, or how fast it writes. I also left out the quotes from Tencent's partners that I could not check against anything outside Tencent's page, apart from the IRAB score, which Tencent reports as the partner's own test.

What I did run is real: the listing checks and the cost and memory arithmetic. The agent task sizes are assumptions, and the memory numbers describe the preview, not Hy4; both are stated where they are used.