Tencent Hunyuan 4: A Multimodal Agent for Blender, Office, and Finance
Tencent's Hunyuan 4 adds vision, 1M-token context, and computer-use skills for Blender, CAD, and office document work. Here's what it can do.
What is Tencent Hunyuan 4?
Hunyuan 4 is Tencent’s newest open weight large language model, the full release following the earlier Hunyuan 4 preview. It is the first version in the line that can see images, not just read and write text, and it holds up to a 1 million token context window, enough to keep an entire project’s files in memory at once. Tencent has said it will publish the model weights and a full system card, though as of this release the architecture details (parameter count, quantization options, exact training data mix) were not yet public.
The headline change from the preview is agentic capability. Hunyuan 4 can operate a computer through both command-line and graphical interfaces, drive tools like Blender and CAD software, and work inside office formats like Word, Excel, and PowerPoint. Tencent’s own blog post says the model was specifically trained on Blender and CAD workflows and adapted for Godot, positioning it as a model meant to act, not just converse.
TL;DR
- Vision support is new to this generation: Hunyuan 4 is the first Hunyuan model that can read images and combine that with instruction-following over long, multi-step tasks.
- Context length reaches 1 million tokens, letting the model hold a full set of project files, financial documents, or code in a single working session instead of chunking them.
- The model was trained for computer use, including driving Blender and CAD directly, which independent testing showed could take a hand-drawn sketch and turn it into a rendered 3D scene with lighting and animation.
- Tencent positions it for real office and finance work, citing strengths in deep research, spreadsheet and document analysis, and multi-file reasoning across Word, Excel, and PowerPoint.
- Benchmark charts from Tencent place Hunyuan 4 ahead of GLM and roughly even with DeepSeek on many tasks, while Claude Opus, GPT, and Kimi models still lead on the hardest coding and agentic tests.
- The jump from the Hunyuan 4 preview to this full release is large, with Tencent’s own tables showing substantial gains across nearly every coding and agentic benchmark it reports.
- Weights and a full system card are expected soon, which is when the exact architecture, parameter count, and training details should become public.
- 1Secure
- 2Regulation-compliant
- 3Auditable
- 4Scalable
How does Hunyuan 4 handle Blender and CAD tasks?
Tencent built Hunyuan 4 with explicit training on Blender and CAD, and the model behaves like an agent that writes and executes code rather than one that just describes a scene. In a hands-on test, the model was given only a hand-drawn sketch of a doner kebab stand and a general instruction, with no detailed spec. It worked through the problem by writing Blender Python code, rendering intermediate results, inspecting its own output, and then correcting specific elements, adjusting the cone shape, the heater grid, and the flame effects, before delivering finished files.
The final output included a complete 3D scene with modeled objects, lighting, and animation: a rotating cone, flickering flames, rising sparks, and falling meat slices rendered as a short looping video, plus a static image and the underlying project files. The result wasn’t flawless (the knife placement relative to the meat was slightly off, and the meat geometry looked more like stacked rings than individual slices), but the model produced a full working 3D asset pipeline from a rough sketch with minimal guidance, which is a meaningfully different task than text-only code generation.
Can it actually do office and finance work?
Tencent markets Hunyuan 4 heavily around office productivity and financial analysis, and a structured test aimed at that claim showed the model handling multi-file, multi-agent reasoning reasonably well. The test setup involved a fictional company’s IPO materials: an Excel file of financials, a CSV of customer data, and a Word memo, with a deliberate numerical mismatch planted between the memo and the spreadsheet. The model was asked to split the reading work across parallel sub-agents, one per file, then combine the findings into a single IPO report.
Hunyuan 4 correctly identified the planted inconsistency between the memo and the spreadsheet, refused to accept the memo’s claims without verification, and flagged which figures needed independent confirmation. It produced a structured report covering margins, debt, customer concentration, and valuation, finishing with a “not yet” verdict on the IPO and listing seven specific conditions that would need to be met first. The full multi-file analysis and report generation completed in under ten minutes.
A separate test asked the model to merge a fake calendar, email thread, and chat log into a single consolidated brief, then translate that brief into more than 80 languages. The model correctly caught planted errors (duplicate entries and a scheduling conflict) in the English version and produced translations across a wide range of languages including Hindi, Korean, Russian, Arabic, Indonesian, and German, with mostly solid results, though a few languages (Urdu in particular) showed rendering or accuracy issues.
How does Hunyuan 4 compare to GLM, DeepSeek, Claude, and GPT?
Nothing gets lost.
Tencent published benchmark charts comparing Hunyuan 4 against competing open and closed models. On most tests, Hunyuan 4 lands in the middle of the pack: ahead of GLM and DeepSeek on many measures, but behind Kimi, GPT, and Claude on a majority of tasks. The model performs closer to the top tier specifically on computer-use benchmarks, web browsing tasks, and an agentic exam, where it sits near several of the larger proprietary models.
The widest performance gap shows up on harder coding tasks and long-context coding benchmarks, a gap Tencent itself acknowledges in its own materials, noting that proprietary frontier models still lead in that category. Compared against its own preview version, Hunyuan 4’s gains are large and consistent across nearly every coding and agentic benchmark reported, suggesting the bulk of Tencent’s post-preview tuning focused on exactly these capabilities.
It’s worth noting that most of these comparison numbers come from Tencent’s own internal testing rather than a neutral third-party benchmark suite, so they should be read as the company’s self-reported view of its standing rather than an independent ranking.
Is Hunyuan 4 worth using for real projects?
For teams doing 3D asset work, document-heavy analysis, or multi-language office automation, Hunyuan 4 looks genuinely more capable than its preview version, based on the agentic tasks it was shown handling: turning sketches into full Blender scenes, catching planted errors in financial documents, and producing structured reports with minimal hand-holding. The 1 million token context window and native vision support address two real limitations of earlier open models in this class.
Where it’s weaker is on the hardest coding and long-context coding benchmarks, where Tencent’s own data shows proprietary models like Claude Opus and GPT still ahead. Teams whose primary need is advanced code generation on difficult problems may still want to default to those closed models, while teams needing an open-weight model that can drive real software (Blender, CAD, office suites) and handle long documents may find Hunyuan 4 closes much of that gap. The caveat is that the full weights and system card weren’t yet available at time of testing, so independent verification of architecture claims and reproducible benchmarks is still pending.
Frequently Asked Questions
What makes Hunyuan 4 different from the Hunyuan 4 preview?
The full release adds image understanding for the first time, extends context length to around 1 million tokens, and shows large benchmark improvements on coding and agentic tasks compared to the preview version, according to Tencent’s own comparison tables.
Can Hunyuan 4 actually control software like Blender?
Yes. Tencent trained it specifically on Blender and CAD workflows, and testing showed it writing and executing Blender Python code, rendering results, checking its own output against a reference image, and fixing specific elements before delivering finished 3D files.
Is Hunyuan 4 open source?
Tencent describes it as an open weight model and has said the weights and a full system card will be released, though the architecture details weren’t fully public at the time of this writing.
How does Hunyuan 4 perform against Claude and GPT models?
On most of Tencent’s own benchmarks, Hunyuan 4 trails Claude and GPT models, particularly on hard coding and long-context coding tasks, but it performs competitively on computer-use, web browsing, and agentic exam benchmarks.
What is the context window size for Hunyuan 4?
Tencent states the model supports up to 1 million tokens of context, which allows it to hold large sets of project files, spreadsheets, or code in memory during a single task.
