Hunyuan 4 Tested: Blender Automation and Real IPO Report Generation
Hands-on tests of Tencent's Hunyuan 4 building a Blender scene from a sketch and writing a full IPO research report from messy files.
What is Hunyuan 4 and what’s new in this release?
Hunyuan 4 is Tencent’s newest open weight foundation model, released as the official follow-up to the Hunyuan 4 preview. The headline upgrade is vision: this is the first version in the line that can see images, not just read and write text. Tencent also says it follows instructions more reliably, holds a 1 million token context window, and can work through long multi-step tasks without losing track of earlier steps. It’s built to operate a computer directly, running command-line tools and graphical apps, including 3D software like Blender and CAD programs. Tencent positions it for building actual front ends and games, with support for engines like Unreal, Unity, and Godot, plus deep research, finance workflows, and editing Word, Excel, and PowerPoint files. The weights haven’t shipped yet at the time of testing; Tencent says they’re coming along with a full system card that will detail the architecture.
TL;DR
- Hunyuan 4 can see images for the first time, a capability Tencent added on top of the preview version, and it’s paired with a 1 million token context window for holding long projects in memory.
- In a hands-on test, the model took a hand-drawn sketch of a doner kebab and drove Blender itself, writing code, rendering, checking its own output against the sketch, and fixing details like flame shapes and the heater grate.
- The final Blender output included a rendered MP4 animation with a turning cone, flickering flames, rising sparks, and falling meat slices, plus a static image and project files, all generated without manual 3D modeling.
- In a simulated IPO due diligence task, the model split reading across parallel sub-agents, caught a deliberately planted numerical mismatch between a memo and a spreadsheet, and delivered a verdict of “not yet” with seven specific conditions.
- On a translation task, Hunyuan 4 converted a mixed bundle of chats, emails, and a calendar into 82 languages, correctly preserving dates, amounts, and merging duplicate entries, with only minor issues spotted in languages like Urdu.
- On Tencent’s own benchmark charts, Hunyuan 4 generally sits ahead of GLM and DeepSeek but behind Kimi, GPT, and Claude on most tests, with the biggest gap showing up on hard, long-context coding tasks.
- The jump from the Hunyuan 4 preview to this full release is described as large across nearly every coding and agentic benchmark Tencent reports.
- 1Secure
- 2Regulation-compliant
- 3Auditable
- 4Scalable
How did Hunyuan 4 perform on the Blender test?
The test gave Hunyuan 4 a single hand-drawn sketch of a doner kebab setup, including a rotating cone of meat, a heater, flames, a knife, and side items like tomato and onion rings. The only instruction was to build a 3D animated scene matching that sketch by driving Blender directly, with no detailed spec or step-by-step guidance.
The model worked through the task by reasoning step by step, writing Blender code, rendering intermediate results, looking at its own output, and correcting mistakes, adjustments to the knife, the flame geometry, and the heater grid were all made autonomously. After roughly an hour, it delivered three outputs: the Blender project file, a rendered MP4 animation, and a static still image.
The animation showed a working 3-second loop: the cone rotating, flames flickering, sparks rising, and meat slices falling. The knife’s position relative to the meat wasn’t perfectly matched, but the overall object placement, lighting, and animation sequencing were handled correctly, including from multiple camera angles. The static image matched the sketch closely, with the heater grate, flames, knife, plate, tomato, onion rings, and sauce bottle all present, though the meat looked more like stacked rings than the hand-drawn version. For a result generated without manual 3D modeling input, the fidelity to the original sketch was notably high.
Can Hunyuan 4 actually handle financial and office work?
The second test simulated a real finance workflow: a fictional Australian data center company considering an IPO, represented by an Excel file of financials, a CSV of customer data, and a Word memo, all internally consistent except for one deliberately planted mismatch between the memo and the spreadsheet.
The prompt asked Hunyuan 4 to split the reading task across three parallel sub-agents, one per file, then combine the findings into a full IPO report covering financial analysis, risk, valuation, cross-document conflicts, and a listing recommendation.
The model launched the sub-agents, correctly found and parsed all three files, and produced a complete Word report inside about ten minutes. Every financial figure in the report matched manual calculations done ahead of time. The report included margins, debt levels, customer concentration, and valuation analysis in tabular form, caught the planted mismatch between the memo and the spreadsheet, and declined to accept the memo’s claims without evidence. The final verdict was “not yet,” accompanied by seven specific conditions that would need to be met before proceeding with a listing. This lines up with Tencent’s stated focus on deep research and finance document handling as a core use case for the model.
How well does it handle multilingual translation?
The third test mirrored a demo scenario Tencent used in their own blog: converting a mix of chats, emails, and a calendar into a single consolidated document, then translating that document into 82 languages, including some less commonly tested ones.
The English baseline output correctly preserved all planted details, dates, amounts, and times, merged duplicate deposit entries, and flagged a scheduling clash between a dentist appointment and a client call. Spot checks across Chinese, Hindi, Korean, Urdu, Spanish, Russian, Indonesian, Persian, Arabic, Swedish, and German showed generally strong output, with Urdu flagged for font rendering issues and a few smaller translation errors noticeable on inspection. Given the breadth of languages covered in a single pass, the overall translation quality held up well, suggesting the model’s multilingual range extends meaningfully beyond major world languages.
Is Hunyuan 4 worth paying attention to against other models?
Tencent’s own benchmark charts place Hunyuan 4 in the middle of the pack on most tests, typically ahead of GLM and DeepSeek but behind Kimi, GPT, and Claude. It performs closer to the top tier on computer use, web browsing, and agentic exam benchmarks, sitting near several larger proprietary models in those categories specifically. The widest performance gap shows up on hard coding and long-context coding tasks, which Tencent itself acknowledges as an area where a gap to top proprietary models remains.
Compared to its own preview version, Tencent’s data shows a large jump across nearly every coding and agentic benchmark, suggesting substantial internal improvement between releases. Against other open models, it trades places with DeepSeek and generally outperforms GLM, while Claude Opus and GPT’s most capable tiers still lead on the hardest tasks. It’s worth noting these are Tencent’s own internal benchmark numbers, not independent third-party evaluations, so they should be read as the company’s framing rather than a neutral ranking. The hands-on tests, image-to-Blender automation, multi-document financial analysis, and large-scale translation, offer a useful real-world complement to those numbers, and in each case the model completed the task with a high degree of accuracy and minimal hand-holding.
Frequently Asked Questions
What is Hunyuan 4?
Hunyuan 4 is Tencent’s latest open weight foundation model, the official release following the Hunyuan 4 preview. It adds image understanding for the first time, supports a 1 million token context window, and is designed for agentic tasks like operating a computer, writing code, and handling office and finance documents.
Can Hunyuan 4 actually control software like Blender?
Yes. In testing, it took a hand-drawn sketch and generated a complete Blender scene, including modeling, lighting, and animation, by writing and executing code, then checking its own rendered output against the original sketch and correcting errors.
Are the weights for Hunyuan 4 available yet?
At the time of testing, Tencent had announced the release but had not yet published the weights or the accompanying system card, which is expected to detail the model’s architecture.
How does Hunyuan 4 compare to GPT and Claude?
On Tencent’s own benchmarks, Hunyuan 4 generally trails the top tiers of GPT and Claude, particularly on hard coding and long-context tasks, but performs competitively against other open models like GLM and DeepSeek, and scores well on agentic and computer-use benchmarks.
What kinds of tasks is Hunyuan 4 best suited for?
Based on both Tencent’s stated design goals and hands-on testing, it performs strongly on 3D and CAD-related automation, multi-document financial analysis and report generation, and large-scale multilingual translation tasks.