Google just released Gemini 3.8 Flash, and we spent some time going through the release notes, API documentation, and demos to see how much of it is relevant to 3D work.
There are a few parts that stand out for artists working in Blender and similar tools. Gemini 3.8 Flash can work with code, images, video, files, and long amounts of context, and Google is putting more emphasis on agents that can use tools and keep working through longer tasks.
We recently looked at a similar direction in our article on GPT-6 Astra and 3D workflows. Gemini is approaching some of the same problems from Google's side.
What is Gemini 3.8 Flash?
Gemini 3.8 Flash is the latest version of Google's Flash model family. Flash models are generally positioned as faster models that can be used more frequently and at larger scale than the company's biggest reasoning models.
According to Google's Gemini 3.8 Flash API documentation, the model supports text, images, audio, video, PDFs, and code. It also has a context window of up to one million tokens.
The same API supports function calling, code execution, file search, search grounding, URL context, structured outputs, and computer use. Those features matter quite a bit once we start thinking about how an AI model could work with a real 3D project instead of answering isolated questions about it.
A Blender project can easily include scripts, reference images, add-on documentation, texture files, renders, and a scene with hundreds or thousands of objects. Giving a model access to more of that information at once makes it easier to imagine workflows where it can understand the project as a whole.
Google has already shown Gemini working with 3D
Google included a few 3D examples in its Gemini 3.8 Flash demonstrations.
One demo has Gemini create an interactive hardware visualizer using Three.js. The viewer generates proportioned 3D device models that can be rotated, inspected, and exploded into separate layers.
Another example uses Google Antigravity to build a small 3D castle game with puzzles, environmental details, and generated textures.
These demos are much smaller than the kind of scenes most professional 3D artists work with, and they're running in web-based environments rather than Blender. Still, they give us a useful example of the model working with code and spatial information at the same time.
That combination is useful because a lot of 3D automation is already possible through scripts. Blender has an extensive Python API, and many repetitive scene operations can already be controlled programmatically. The harder part is giving a system enough context to understand what needs to happen and whether the result is correct.
How Gemini 3.8 Flash could fit into Blender
Google hasn't announced an official Gemini 3.8 Flash integration for Blender.
There are already several ways developers can connect AI models to Blender, though. Blender exposes a large portion of its functionality through Python, and Gemini supports function calling, code generation, external tools, and computer use.
A practical use might be cleaning up an imported scene. An agent could inspect object names, find missing materials, check texture paths, organize collections, look for obviously incorrect scale, or write a Blender script for operations that need to be repeated across hundreds of objects.
Most of those tasks are already possible with Blender's existing tools. The difference is that today we usually have to decide exactly what operation to run and how to run it. An agent could potentially handle more of that interpretation.
For artists who regularly work with imported assets, this is also close to what we're already trying to reduce with features such as one-click material import. There are plenty of small setup steps in a 3D workflow that don't need much creative judgment but still take time.
Materials are a straightforward example
Materials give us a pretty clear example because the workflow involves several predictable steps.
Say we need a weathered stone floor. We can generate the surface using an AI PBR material generator, download the maps, bring them into Blender, create the shader, assign the correct color spaces, connect the normal and displacement maps, set the scale, and then render the result.
If we're using materials from a larger PBR material library, the generation step may disappear, but most of the setup inside the 3D application is still there.
An agent with access to both systems could eventually handle more of that setup. It could request the material, retrieve the PBR maps, build the Blender shader, apply it to the right object, set a starting texture scale, and render a preview.
Google hasn't demonstrated that specific workflow with Gemini 3.8 Flash, so we wouldn't treat it as a current Gemini feature. It is simply the kind of workflow that becomes possible when a model can combine tool calls, code, files, and visual input.
The same setup could work with seamless materials, where the agent wouldn't need to solve tiling problems before the material is applied to a larger surface.
Images and video add another useful layer
Gemini 3.8 Flash can take both images and video as input. That matters in 3D because a lot of the feedback we use while working isn't text.
We look at the viewport. We compare the render against a reference. We scrub through an animation to find where something breaks. We notice that a normal map is inverted or that a material scale is obviously wrong.
Google has also been working on agentic video understanding, where Gemini can use tools to inspect relevant parts of a video while answering a question instead of processing the entire video in the same way.
A version of that workflow could be useful with viewport recordings. You could give the model a recording of an animation problem along with the relevant project information and let it work out where the issue is likely coming from.
The model would still need access to Blender or another tool to make the actual change, but visual input gives it more of the same information we use when diagnosing a scene ourselves.
Computer use could cover the parts scripts don't
A lot of Blender can be controlled through Python, but not every workflow starts with a clean API call.
Gemini 3.8 Flash includes computer use support in preview. That gives developers another option for tasks that still require navigating a visual interface.
We wouldn't expect computer use alone to be a good way to automate a complicated Blender project. Clicking through a dense 3D interface is slower and less reliable than calling an API or running a script when those options are available.
A more realistic setup would probably use several approaches together. Scripts would handle predictable operations, tool calls would connect external services, and computer use could deal with the occasional part of the workflow that doesn't expose a cleaner interface.
Gemini 3.8 Flash Cyber is a separate model
Google released Gemini 3.8 Flash Cyber alongside the standard Flash model.
Flash Cyber is focused on security work, particularly finding software vulnerabilities and generating patches. Google says its Chrome security team found that the model produced 2.6 times more correct vulnerability patches than the larger commercial models they tested.
The model is currently available only to selected defenders through Google's Fairwind Program, so it isn't directly relevant to most 3D artists.
It does show how Google is starting to split some agent work into more specialized models. Over time, we could see the same thing happen in creative software, with different models handling code, visual checking, asset generation, or other narrow jobs rather than one model trying to handle an entire production pipeline by itself.
What Gemini 3.8 Flash means for 3D workflows
We don't expect people to replace Blender with Gemini. Blender already has decades of work behind its modeling, shading, animation, rendering, and scripting systems. The same is true for Houdini, Substance, renderers, asset libraries, and the other software most 3D artists use.
There is more room to improve the way those tools connect.
A material might come from twirl, an asset might come from a library, the scene might live in Blender, and a custom script might handle part of the setup. Right now, we usually move the files and settings between those systems ourselves.
Models like Gemini 3.8 Flash give developers more ways to automate that middle layer. Tool calling can connect services, code can control applications, visual input can help inspect results, and longer context lets the model keep more of the project in view while it works.
Google's current 3D demos are still fairly simple compared with production work, so there is a lot left to prove. But the direction is clear enough to start building around. For 3D artists, the useful version of AI may end up being less about generating one finished asset and more about helping with all the small steps that happen around it.