--- title: "How Good Is Tencent Hy4 at Long Agentic Tasks? Hands-On Tests" url: "https://goremy.ai/blog/articles/tencent-hy4-hands-on-test-blender-finance" author: Luis Chavez-Mattos editor: Luis Chavez-Mattos editor_title: Director, Product published: "2026-11-05T20:00:15.006Z" updated: "2026-11-05T20:00:15.006Z" topic: Open-Weight Model Benchmarks description: "Tencent's Hy4 was tested on a sketch-to-Blender scene, a multi-agent IPO report with a planted error, and 82-language output. Results and caveats." --- # How Good Is Tencent Hy4 at Long Agentic Tasks? Hands-On Tests > Tencent's Hy4 was tested on a sketch-to-Blender scene, a multi-agent IPO report with a planted error, and 82-language output. Results and caveats. Edited by Luis Chavez-Mattos, Director, Product ยท Published 2026-11-05 ## How good is Tencent Hy4 at long multi-step tasks? Good, but not the best. In three hands-on agentic tests run through the Hermes agent, Hy4 built a Blender scene from a sketch, produced an accurate IPO report that caught a planted data conflict, and generated a working 82-language offline page. Tencent's own charts place it ahead of GLM and DeepSeek on most rows, but behind Kimi, GPT and Claude, with the widest gap on hard, long coding tasks. ## TL;DR - Hy4 is the official release that follows the Hy4 preview. It is Tencent's newest **open-weight model**, and it now accepts images as well as text. - In a sketch-to-3D test, the model drove Blender through code, rendered its own output, and **fixed its own mistakes** over roughly an hour. - In a finance test, three parallel sub-agents produced an IPO report in under 10 minutes. Every figure matched the tester's own calculations, and the model **caught the planted mismatch** between a memo and a spreadsheet. - In a multilingual test, Hy4 produced a page covering 82 languages with correct dates, amounts and times in English, though the tester spotted some mistakes and a font problem in Urdu. - Tencent's charts put Hy4 **mid-pack among frontier models**, ahead of GLM 5.3 on most rows and trading places with DeepSeek. Claude and GPT still lead on harder tests. - Many benchmarks are Tencent's internal ones, and the final weights and system card had not yet been released at the time of testing. ## What is Hy4 and what changed from the preview? Hy4 is Tencent's open-weight flagship, released as the official version after Hy4 preview. According to the announcement, it sees images for the first time, follows instructions better, holds long context, and finishes longer multi-step jobs. It supports a 1 million token context window and can operate a computer through the command line and graphical apps, including Blender and CAD tools. Tencent also says it is tuned for front-end and game work with engines such as Unreal, Unity and Godot, and for deep research, finance work, and editing Word and PowerPoint files. Architecture details for the final model are still pending. Tencent said the weights will arrive with a full system card. The Hy4 preview card does give specifics, and they are a reasonable guide to the family: | Property | Hy4 preview | |:---|:---| | Architecture | Mixture-of-Experts | | Total parameters | 770B | | Activated per token | 49B | | Layers | 78 | | Routed experts | 256 (top 8 active, plus 1 shared) | | Attention | Gated DeepSeek Sparse Attention | | Context length | 1M | | License | Apache 2.0 | Treat these as preview numbers. The final release may differ. ## How did Hy4 do on the sketch-to-Blender test? The brief was a hand-drawn sketch of a doner kebab stand. The goal given to Hy4 inside Hermes agent was to build a 3D animation scene by driving Blender, with little else specified. The tester also asked for an MP4 and a still image. The run took about an hour. The visible reasoning showed the model writing Blender code, rendering, inspecting its own picture, and then correcting parts such as the knife, the flames and the heater grate. It delivered all three requested files and checked them against the sketch. The result had the full set of parts: meat cone, heater, flames, tomato, onion rings, plate and sauce bottle, plus lights and animation. The render was a 3-second loop with the cone turning, flames flickering, sparks rising and slices falling. Flaws were small. The knife sat slightly away from the meat, and the meat looked like stacked rings, less realistic than the drawing. That self-inspect-and-fix loop is the interesting part. The model used its new vision to judge its own output rather than shipping the first render. ## Can Hy4 handle a multi-file finance report with parallel sub-agents? This test used a fictional Australian data center company weighing an IPO. The inputs were an Excel file of financials, a CSV of customers and a Word board memo. The prompt told Hy4 to split the reading across three parallel sub-agents and combine their findings into an IPO report. One discrepancy between the memo and the spreadsheet was planted on purpose, and the tester already knew the correct numbers. Hy4 launched the sub-agents, found the files, and used locally installed tools without hand-holding. It returned a full Word report in under 10 minutes, with appendices. Findings: - Every number matched the tester's own calculations. - Margins, debt, customer concentration and valuation were covered, mostly in tables. - It caught the planted capacity mismatch between memo and spreadsheet. - It declined to accept the memo's claims at face value and flagged what needed proof. - Its verdict was "not yet," with seven specific conditions to meet first. Tool calls were described as targeted. This is a small controlled test with a single planted error, so it shows competence at cross-file checking, not a guarantee on real filings. ## How does Hy4 perform on 82-language output? Tencent's own demo turns chats, emails and a calendar into one offline page. The tester built fake versions of those inputs and asked Hy4 to translate the whole board into 82 languages, including two unexpected ones. The English version had all five planted items right, with correct dates, amounts and times. Duplicate deposits were merged, and a planted clash between a dentist appointment and a client call was flagged. The page had a language drop-down. The tester spot-checked Chinese, Hindi, Korean, Urdu, Spanish, Russian, Indonesian, Persian, Arabic, Swedish and German. Urdu showed a poor font and a few visible errors. The tester could not verify most of the 82 languages, so quality across the long tail is unconfirmed. ## How does Hy4 compare with GLM, DeepSeek, Kimi and closed models? Tencent published charts, and the reviewer's reading of them was consistent: | Comparison | Where Hy4 lands | |:---|:---| | GLM 5.3 | Ahead on most rows | | DeepSeek | Competitive, trading places | | Kimi | Behind on most tests | | GPT and Claude | Behind on most tests, especially harder ones | | Hy4 preview | Large jump across most coding and agentic tests | Hy4 gets close to the top on computer use, web browsing and an agentic exam. The widest gap is on a hard, long coding test, which fits Tencent's own note that a gap to the top proprietary models remains. The video named Claude Opus 5.5 and GPT 6.1 as the models still leading on the harder rows. The preview card adds a human evaluation. 163 internal experts rated outputs on 203 engineering tasks. Hy4 preview edged GLM 5.3 (2.99 vs 2.92 average) and Kimi K3 (2.99 vs 2.94). Win rates were 46.8% against GLM and 51.2% against Kimi, with 40.4% and 40.9% losses. That is Tencent's own panel, and the margins are slim. ## Should you trust these results? Read them with three caveats. 1. **Internal benchmarks.** Many of the tests are Tencent's own, so the numbers show Tencent's view, not a neutral ruling. 2. **Small tests.** Three tasks with one planted error show capability, not reliability at scale. 3. **Known preview issues.** Tencent's preview card admits the model can reason longer than needed and over-verify its own work. The hour-long Blender run is consistent with that, though the sources don't say whether the final release fixed it. The most useful approach is to run your own task through your own agent harness and compare against the published numbers. ## Frequently Asked Questions ### Is Hy4 open weight? Tencent describes Hy4 as open weight and said the weights and a full system card are coming. The earlier Hy4 preview weights, including an FP8 version, are already on Hugging Face under Apache 2.0. The license for the final release should be confirmed when it ships. ### Can Hy4 read images? Yes. The official release is the first Hy4 version that can see as well as read. In the Blender test it used that ability to look at its own renders and compare them with the original sketch. ### How big is the Hy4 context window? The context window is 1 million tokens, which Tencent frames as enough to hold a whole project. The preview card lists the same 1M context length. ### Is Hy4 better than GLM 5.3 and DeepSeek? On Tencent's charts, it is ahead of GLM 5.3 on most rows and roughly level with DeepSeek. Tencent's preview human evaluation also had it slightly ahead of GLM 5.3 and Kimi K3. Claude and GPT still lead on harder tests. ### How do you run Hy4 locally? The final weights were not yet out at the time of testing. The preview card lists a 770B-parameter Mixture-of-Experts model with 49B activated per token, so expect a multi-GPU server rather than a single consumer card.