---
title: "Tencent Hy4 launches through hosted services, with weights still to come"
description: "Tencent Hy4 access and benchmark evidence, with practical checks for visual tasks, computer use and deployment decisions."
url: "https://promptszone.com/tencent-hy4-hosted-release-weights-benchmark-caveats"
published: "2026-11-04T15:00:05Z"
categories: ["New Models"]
---

# Tencent Hy4 launches through hosted services, with weights still to come

Hy4’s access options and a cautious reading of Tencent’s finance-task comparison.

![Four-model chart of Tencent-reported finance-task scores: Hy4 3.04, DeepSeek 2.97, Kimi and GLM 2.68. Not an independent overall ranking.](https://promptszone.com/assets/uploads/8282b5e0a46869ab86ce2eeee4fa8a62.png)

**Summary:**

A practical look at access, evaluation limits and what to test before adopting Hy4.

Tencent [announced Hy4 on November 4](https://en.prnasia.com/releases/apac/tencent-launches-hy4-joining-the-ranks-of-the-world-s-leading-open-source-ai-models-554689.shtml), adding visual understanding and computer use. Hosted access and downloadable weights are different milestones: Tencent says weights and system cards will follow in the coming days.

## Where to try it

Tencent lists TokenHub, WorkBuddy, CodeBuddy, Yuanbao and ima among its access channels. Product-specific eligibility, regional availability and account requirements still need checking; we have not tested access from every market.

[OpenRouter lists tencent/hy4](https://openrouter.ai/tencent/hy4) at $0.417 per million input tokens, $1.251 per million output tokens and $0.021 per million cache-read tokens, checked November 4. These are that provider’s displayed rates, not a universal Tencent price or a guaranteed future price.

## One comparison, not an overall ranking

Tencent reports a developer-led blind evaluation with 213 financial professionals across 579 tasks. Its reported average scores are shown below; higher is better.

| Evaluation | Hy4 | DeepSeek-V4.1-Flash-1017 | Kimi K3.1 Pro | GLM-5.3 |
| --- | --- | --- | --- | --- |
| Finance tasks, reported average | **3.04** | 2.97 | 2.68 | 2.68 |

Source: Tencent’s November 4 announcement. This is not an independent test. We could not establish enough comparable multi-benchmark data to construct a fair broader scorecard. The release does not give sufficient detail here to reproduce the test or assess statistical uncertainty. A small numerical lead should not be treated as proof that one model is best for your workload.

## How to evaluate it for your work

**Our suggested test:** keep the prompt, input documents, tools and success criteria fixed across candidates. Evaluate whether the answer cites the right evidence and whether a tool action actually produces the intended result, not simply whether the explanation sounds confident.

- **For visual tasks:** use documents you may share, include difficult chart labels and ask the model to identify uncertain readings. Check the output against the original.
- **For computer use:** begin in a disposable environment with no credentials or live financial transactions. Review the proposed actions and verify the resulting files.
- **For deployment:** compare latency, retries, total task cost and failure recovery. Token prices alone do not measure the cost of a successfully completed job.
- **For self-hosting:** wait for the final weights, license and deployment documentation. Do not substitute an older preview checkpoint while claiming to test this release.

Those checks help turn a launch announcement into a practical decision. For another example of separating an API launch from pending downloads, see our [LongCat-2.5 access and benchmark coverage](https://promptszone.com/meituan-longcat-2-5-api-release-benchmark-comparison).

## Image and reporting notes

The featured chart is an AI-assisted visualization of the cited values, checked against the source. It uses Tencent’s official provider logo as a labeled fallback, not a distinct Hy4 model logo. Logo source: [Tencent’s editorial media library](https://www.tencent.com/newsroom/media-resources/). No endorsement is implied. We did not independently benchmark the model. Reporting checked November 4, 2026.