Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Tencent Hy4: vision and computer use, weights pending

Written November 5, 2026

Hy4 adds visual understanding and computer use to Tencent's flagship, at $0.83/$2.50 per million tokens. The weights are promised, not published.

What it is

Tencent released the full version of Hy4 on November 4, ten weeks after the text-only Hy4 preview of August 28. The full release adds two things for the first time in the Hy line: visual understanding and computer use. Hy4 can read fine details in images, analyze complex charts and documents, and reason across multiple images. It can also operate command-line tools and graphical interfaces, chaining steps across applications. Tencent says it trained the model on professional software such as Blender and CAD, with workflows spanning several programs, and that development focused on office tasks, financial analysis, and game development. In one demonstration, the model found a roof component about two metres across in a 2.8-gigapixel aerial photograph.

The size figures in circulation — 770 billion total parameters with 49 billion active, in a mixture-of-experts design, with a 1-million-token context window — come from the August preview's model card. Tencent's November announcement did not restate them, so treat them as carried-over preview figures rather than freshly confirmed final specs. The full model's own system card has not been published yet.

Where it runs, and what it costs

Hy4 is available now through APIs only. It runs on Tencent Cloud's TokenHub, on OpenRouter as tencent/hy4 with a 1,048,576-token context window, and inside Tencent's own WorkBuddy, CodeBuddy, Yuanbao, and ima products. WorkBuddy and CodeBuddy offer a limited free usage quota from November 4 to November 15.

Standard API pricing is unchanged from the Hy4 preview: $0.834 per million input tokens and $2.501 per million output tokens, with cached input at $0.042, per TechNode's report of the launch and the OpenRouter listing. Two discounts apply at the moment: OpenRouter currently shows the model at 50% off, $0.417 in and $1.251 out, and Tencent Cloud TokenHub carries a 50% discount through December 31. Both are time-limited, so the price to budget on is the standard $0.834/$2.501, not the discounted rate.

What is still a promise

The weights and the system card are promised, not published. Tencent says it plans to release the model weights and a comprehensive system card shortly, with support from the vLLM and SGLang communities, but has given no date. The August preview shipped its weights on day one under Apache 2.0; the full model has no published weights, licence, or model card yet, so do not assume the preview's licence carries over. A lower-cost Hy4 Flash model is also planned, with no date given. Until the weights land, Hy4 is an API-only model: usable through Tencent Cloud, OpenRouter, and Tencent's own products, and not downloadable.

The contrast with this week's other launches is the point. Kimi K3.1 Pro's weights landed on the promised date, and Ripple-65B shipped its weights on day one. Hy4's weights are a promise with no date, so any plan that depends on self-hosting Hy4 is a plan that depends on a future release.

The benchmarks, read as vendor claims

Every score here is Tencent's own, and none has been independently rerun. On OSWorld 2.1, a test of operating a computer, Tencent reports 78.9 for Hy4, against 81.8 for Claude Opus 5.5, 79.6 for Kimi K3.1 Pro, and 75.2 for GPT-6.1 Sol. On Terminal-Bench 4.0, Hy4 scores 44.4 against 66.4 for Opus 5.5. On a finance evaluation of 579 real-world tasks reviewed by professionals, Tencent says Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash at 2.97. Tencent's own announcements give different figures for some of its internal measurements, so treat every vendor number here as company-reported, not independently checked. Tencent itself says a notable gap to the top closed models remains, so test before switching.

What to do

  1. Try it through the API first. Use the free WorkBuddy and CodeBuddy quota before November 15, or the discounted TokenHub rate before December 31, to test vision and computer use on real tasks.
  2. Budget on the standard price. The 50% discounts end. Plan on $0.834 in and $2.501 out per million tokens.
  3. Do not plan a self-hosted deployment yet. No weights, no licence, no date. If the work needs downloadable weights, use a model whose weights are already published.
  4. Test computer use with guardrails. A model that operates a live desktop needs permissions, logging, and human approval on consequential actions. Test on a machine where a wrong click costs nothing.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Kimi K3.1 Pro weights are out: what 1.56 TB buys · Ripple-65B: the $32M open model topping coding tests