Tencent Hy4 preview — the open 770B MoE whose official pages print three different token ceilings, two reasoning levels, and a 46-row appendix against the generation it replaces

What is Tencent Hy4 preview?

Tencent Hy4 preview is the Mixture-of-Experts flagship that the Tencent Hy Team released and open-sourced on 28 August 2026, and it is shipped as an early version of the Hy4 generation rather than as a finished one — Tencent says so itself and lists two known limitations alongside the release. The backbone is 770B total parameters with 49B activated per token across 78 layers: the first layer is a standard dense feed-forward network, the remaining 77 are mixture-of-experts layers holding 256 routed experts plus one shared expert each, and every token activates the top eight routed experts together with the shared one. A native multi-token-prediction layer of roughly 10B total and 0.7B activated parameters is built in for speculative decoding, so no separate draft model has to be deployed. Attention uses Gated DeepSeek Sparse Attention with IndexCache for cross-layer sparse index reuse, the residual pathway uses identity Hyper-Connections to widen inter-layer information flow, the hidden size is 6,144, the attention head count is 64, and the vocabulary is 120,832 entries. The weights are published under Apache 2.0 in both BF16 and FP8, on Hugging Face, ModelScope, GitCode and CNB, with vLLM and SGLang serving recipes and a prebuilt container image on day one. Three numbers on the official pages describe three different limits rather than one, and keeping them apart is the point of the specifications block below: the context window is 1,024k tokens, the maximum input is 960k tokens, and the maximum output is 64k tokens. Input is text and output is text; the model page and the model list both record function calling, structured output through a response format, and cache support. Thinking is on by default and there are only two settings for it, a full chain-of-thought mode that is the default and a way to turn it off. Tencent also reports its own blind comparison: 163 internal experts rated 203 engineering tasks and put this model at 2.99 out of 4.00 against 2.92 for GLM 5.3 and 2.94 for Kimi K3, figures the company measures itself rather than figures from an independent lab. The identifier is hy4-preview, and the model is listed as available in preview rather than as a finished release.

What Tencent Hy4 preview is worth using for

Official Tencent Hy4 preview specifications

Developer

Tencent Hunyuan

Model ID

hy4-preview

Official context window

1,024k tokens (also written 1M)

Official maximum input

960k tokens

Official maximum output

64k tokens

Official modalities

Text in, text out

Tencent Hy4 preview and Hy3 side by side

Every value below comes from a single page — the model repository README on the Tencent Hunyuan GitHub organisation — and from a single image within it, the benchmark appendix in assets. That appendix prints ten columns across 46 rows, one column per model, and the two columns reproduced here are the previous generation, Hy3, and this page’s model, Tencent Hy4 preview. One official artefact therefore accounts for both columns of every row, which is why the table can be read as a comparison at all rather than as two lists placed next to each other. Read the direction before the size of the gaps: the left column is this page’s model and the right column is the generation it replaces, and the later model scores higher on every one of the 46 rows. The gaps are not uniform, and the two largest carry the most information about what changed. On the internal benchmark named Hy-CompanyBench V2 the older model scores 29.8 against 62.4, and on SWE-Marathon it scores 5.0 against 31.9 — both of which are near-doubling or larger, and both of which sit in the agentic-coding and long-horizon-work groups rather than the pure-reasoning group. The smallest gaps cluster at the other end: GPQA Diamond moves from 90.9 to 92.3, and Hy-BrowseComp-Pro2 from 55.0 to 56.1. Nine rows carry an internal label in their names, which the appendix prints rather than hides. Six further columns exist on the same image — DeepSeek V4 Pro 0813, Qwen 3.8 Max, GLM 5.3, Kimi K3, GPT 5.6 Sol and Claude Opus 5 — and they are deliberately not reproduced, because folding six more models into a two-generation comparison would turn it into something else. Three facts travel with the figures rather than being dropped. The first is provenance: the appendix footnote states that every model is evaluated at its highest available reasoning setting, and that results marked with an asterisk are the vendor’s own testing, so no row here is an independently verified measurement. The second is a warning that applies to the right-hand column specifically: the same footnote states that certain Hy3 scores may differ from previously reported results after updates to the evaluation harness, the judge model and anti-hacking mechanisms, so this column is Hy3 as measured under current conditions rather than Hy3 as published at its own launch. The third is that several rows carry harness notes — SWE-bench Multilingual and SWE-bench Pro use a specific agent scaffold, DeepSWE uses a different one in a sandbox with 8 CPU cores and 16 GB of memory, SWE-Marathon doubles the official agent timeout, and Terminal-Bench 2.1 is run with a Claude Code harness allowing up to 500 turns and a 12-hour timeout — so a row is a number plus a procedure, not a number alone. No row predicts the quality or the speed of the conversation you are about to have.

How to use Tencent Hy4 preview online

1. Choose the model

   This page opens with Tencent Hy4 preview already selected. The model navigation and the selector inside the chat box stay in sync, switching models keeps the current conversation, and starting a new chat is how you get a clean context. Two naming facts are worth fixing before you send anything. The official pages write the model as Hy4 preview in prose and as hy4-preview as the identifier, while this site’s display label, which is what the navigation and the heading show, reads Tencent Hy4 preview — the vendor name attached to the same model. And the model is listed as a preview build rather than a finished release, so treat published behaviour as something that can change.

2. Decide what you expect from thinking, because this box has a switch and no level

   The official pages list thinking as enabled by default and give exactly two settings — a full chain-of-thought mode at high effort and a way to turn it off — and the model-specific table is the one that governs. Other models on the same documentation page carry three or more settings, so it is easy to assume this one does too, and the generic sentence about tool calling on another page does mention a setting this model does not have. What this page gives you is neither the default nor the two values. The catalog marks reasoning as supported but publishes no levels at all, so the switch can be turned on and off and nothing can be chosen, and it starts off rather than on. There is therefore no lever here for how much thinking a task gets, and the only way to shape it is the prompt itself. Note that this is not the same statement as the official ones: the model thinks by default, the API accepts two values, and this box asks for neither unless you flip the switch.

3. Sign in, send, and check the figures against the official pages

   Sending a message needs an account with available points. The composer reads a reference price from the live model catalog and a finished request reports what it actually cost. Two ceilings are set by this site rather than by the model, so plan around them: replies are requested with at most 4,096 output tokens, and no context figure is shown anywhere. Because the rates and the token ceilings quoted on this page are the ones the official pages print, re-check them there before relying on a figure for planning — and treat any third-party summary that presents a relay, resale or converted figure as out of scope, because no official page publishes one. Nothing here converts points into yuan, and nothing on the official pages puts an expiry on the rates.

What Tencent Hy4 preview costs, and why this page carries no end date

The first thing to establish about this model’s price is what is not attached to it. No official page puts an end date on any of its rates, and no official page publishes a promotional or introductory price for the model itself. Both the pricing page and the research page were checked on 5 October 2026. The official rate card lists this model with three figures per million tokens: 6 yuan for input, 18 yuan for output and 0.3 yuan for cache hits, with no input-length tier and no peak-versus-off-peak rule, both of which are printed for other models on the same page as a dash. The research page prints the same three numbers in the same order under its own heading. Read against the previous generation the card is a clear step up: Hy3 is listed at 1 yuan for input, 4 yuan for output and 0.25 yuan for cache hits, so fresh input costs six times as much here and output four and a half times as much, while the cache rate rises by a fifth. That is worth stating plainly because it means the cache is not where the money is saved on a small workload — cache hits are cheap in both, and the cost of a task scales with fresh input and output instead. Two further readings follow from the layout of the table. Every cell in this model’s row is a bare number except the two that are dashes, so there is no conditional pricing to trip over: no length threshold that re-prices the whole request, which the same page does document for another vendor’s models, and no peak-hour multiplier. And the price page is dated while the rate is not: the page carries its own last-updated timestamp, the rate carries nothing, so the honest statement is that the figures were true when the page was last edited, not that they hold until a given date. One more charge is attached to this model by name, and it is not a rate at all: the model list records cache support as a capability, which is what the third figure on the card is for. The model also appears in a separate product announcement that offers a first-month discount for new customers of that product, and because that offer names no window anywhere on the page, this page does not repeat it. Nothing above is this site’s checkout price. This site bills points against a catalog reference price of 148.7676 points per million input tokens, 446.1243 points per million output tokens and 7.4919 points per million cached read tokens, and there is no conversion between points and yuan.

Tencent Hy4 preview FAQ