A flagship that can finally see, and click

Hy4 is the full version of the model Tencent previewed in August, with both training stages scaled up and two capabilities the preview did not have. It can now read images — charts, documents, photographs — and it can operate a computer, through the command line and through on-screen applications. Tencent says the computer training covered professional software such as Blender and CAD tools, and tasks that run long and cross several applications. In one demonstration the model picks out a roof fitting about two meters across in a 2.8-billion-pixel aerial photograph. That is a demonstration, not a measurement, and Tencent offers it as one.

The release is live where Tencent’s models usually land first: Yuanbao, WorkBuddy, CodeBuddy, ima and the WeCom assistant, plus Tencent Cloud’s TokenHub. Developers outside China can call it through OpenRouter, which lists a context window of just over a million tokens and image input, at the preview’s prices, currently displayed at half off.

The limits section is the launch

Most launch posts bury their weaknesses. Tencent’s gives them a section. Hy4, the team writes, may sit among the strongest open models, and it still has “a significant gap” to the top closed models, above all on the hardest internal tests it runs: frontier mathematics, long software projects, model development and visual questions.

Its own chart supports the sentence. On Terminal-Bench 4.0, a test of finishing real tasks in a terminal, Hy4 scores 44.4 against 66.4 for Claude Opus 5.5 and 55.5 for GPT-6.1 Sol. On ProgramBench it scores 30.0 to Claude’s 56.5. The closest result is on OSWorld 2.1, Tencent’s computer-operation benchmark: Hy4 scores 78.9, against 81.8 for Claude and 75.2 for GPT-6.1 Sol. These are Tencent’s reported results, not independently reproduced scores. Against other open models in Tencent’s chart, the picture splits. In Tencent’s own comparisons, Hy4 beats DeepSeek V4.1 Flash and GLM-5.3 on several reported tests, while Kimi K3.1 Pro scores higher on others. The results do not establish an overall ranking among these models. In a finance evaluation organized with Tencent’s own fintech unit, 213 experts assessed 579 real tasks. Tencent reports an average of 3.04 out of 4 for Hy4, against 2.97 for DeepSeek and 2.68 for both Kimi and GLM. The evaluation provides a useful indication of the company’s target workloads, but it is not an independent comparison.

Open, at a later date

The word doing the most work in the release coverage is “open”. What is actually available today is the API. The model with published weights is the August preview, not the full Hy4 described in this release. Today, what is open is the preview: a 770B-parameter mixture-of-experts model, 49B active, published in August under Apache 2.0. For the full Hy4, Tencent says it plans to release the weights, with serving support from the vLLM and SGLang projects, together with a system card carrying fuller safety results. It gives no date, and it has not said which license the full version will carry. Until both appear, Hy4 is an API model with an open ancestor.

On safety, Tencent says it kept defensive security material in training — proof-of-concept generation, bug fixing — and removed material on exploits, reverse engineering and penetration testing, the same trade Moonshot described for Kimi K3.1 Pro’s open release. Its own security labs assessed the offensive cyber risk and its own biology-lab test scored 30.4 out of 100, below the threshold Tencent calls high-risk. Those are the vendor’s labs and the vendor’s threshold, stated here because publishing them at launch is still uncommon among Chinese labs, not because they settle anything.

The practical offer, for now, is price. OpenRouter lists Hy4 at $0.834 per million input tokens and $2.501 per million output tokens. At the time of checking on November 4, both were displayed at 50% off: $0.417 and $1.251 respectively. These are marketplace prices, separate from Tencent Cloud’s TokenHub pricing and promotion. A team whose work looks like OSWorld — long stretches of ordinary computer operation — gets a model within three points of the leader on Tencent’s own test, at API prices. A team whose work looks like Terminal-Bench gets a model 22 points behind, on the same chart. Tencent, to its credit, printed both numbers itself.

Sources

  1. [1] Tencent Hunyuan, “Hy4 正式发布” (Hy4 official release, in Chinese), November 4, 2026 — Tencent’s launch post on hunyuan.tencent.com; the hunyuan.tencent.com page itself returned no extractable text to the Gate 1 fetcher, so its content was verified through the full English rendering of the announcement carried by ChainCatcher and through MadRobot’s first-hand read below: scaled-up pre-training and post-training; visual understanding and computer use for the first time, across the command line and graphical interfaces, with specialized training on professional software such as Blender and CAD and on long tasks spanning several applications; a limits section conceding “a significant gap” to the top closed models; pricing unchanged at ¥6 / ¥18 / ¥0.3 per million tokens with TokenHub calls at half price until December 31, 2026; the full weights and a system card planned with vLLM and SGLang support, with no date and no license stated; Hy4 Flash due soon; defensive security material kept in training and offensive material removed, with the company’s own lab assessments including a biology-lab score of 30.4 out of 100Read source
  2. [2] MadRobot, “Tencent Hy4 is out: it can use a computer but still trails Claude,” November 4, 2026 (read first-hand in full at Gate 1 — the detailed English account of Tencent’s launch post: the benchmark chart, Terminal-Bench 4.0 Hy4 44.4 against Claude Opus 5.5 66.4, GPT-6.1 Sol 55.5 and Kimi K3.1 Pro 66.8, ProgramBench 30.0 against 56.5, OSWorld 2.1 78.9 against Claude 81.8 and GPT-6.1 Sol 75.2; the finance evaluation run with Tencent’s own fintech unit, 213 experts across 579 real tasks, Hy4 averaging 3.04 out of 4 against DeepSeek 2.97 and Kimi and GLM 2.68; the aerial-photograph demonstration; TokenHub calls half price until December 31; as of November 4, only the August preview’s weights were on Tencent’s Hugging Face page)Read source
  3. [3] OpenRouter, “Hy4 — API Pricing & Providers,” re-read first-hand at Gate 1 on November 4, 2026 (release date November 4, 2026; context window 1,048,576 tokens; maximum completion 65,536 tokens; text and image input, text output; tool calling and structured outputs supported; list prices $0.834 input / $2.501 output / $0.042 cached per million tokens, displayed at 50% off at the time of checking)Read source
  4. [4] Hugging Face, tencent/Hy4-preview model card, read first-hand at Gate 1 on November 4, 2026 (the August preview: 770B total parameters, 49B activated per token, weights released under the Apache License 2.0; the full Hy4’s weights were not on Tencent’s Hugging Face page at the time of checking)Read source
  5. [5] WeDoAny, “Tencent Releases Official Version of Hunyuan Hy4 with New Visual Operation Capabilities,” November 4, 2026 (corroborates the product list — TokenHub, WorkBuddy, CodeBuddy, Yuanbao, ima, WeCom assistant — the unchanged preview-level pricing of ¥0.3 cached / ¥6 input / ¥18 output per million tokens, and the plan to open the full version’s model weights and system card afterwards)Read source
  6. [6] PANews, “Tencent Hunyuan Releases Hy4, Its Latest-Generation Large Language Model,” November 4, 2026 (via Jiemian News; first visual understanding, deep reasoning and tool calling; live in the same Tencent products)Read source