Kimi K3.1 Pro claims strong 3D and coding skills in Blender, Unreal, and CAD. Here's what its real demos show, pass and fail.
What is Kimi K3.1 Pro? #
Kimi K3.1 Pro is a large open weights language model from the Chinese AI lab Moonshot AI, built on top of the earlier Kimi K3 released in July. It uses a mixture of experts design with roughly 2.8 trillion total parameters, of which only around 104 billion activate per request, which keeps it fast and relatively cheap to run despite its size. It has a 1 million token context window, understands images as well as text, and Moonshot has marketed it specifically around coding, agentic tasks, finance work, and 3D content creation in tools like Blender, Unreal Engine, Unity, Godot, and CAD software.
TL;DR #
- Moonshot AI positions Kimi K3.1 Pro as a direct competitor to Claude Opus 5.5 and GPT-6 Astra, claiming it matches or beats them on several benchmarks including Humanity’s Last Exam.
- Independent testing tells a more modest story: Artificial Analysis ranks it fourth overall, and Arena’s web dev leaderboard puts it eighth, both strong but clearly behind the top closed models.
- 3D rendering is a real strength , shown in demos like a fully interactive, physics-accurate 3D Rubik’s Cube simulation that correctly scrambles and solves itself, something many open weights models struggle with.
- Complex spatial tasks still trip it up , such as modeling a bowline knot in Blender, where the rope geometry didn’t actually knot correctly despite understanding the general shape.
- Pricing is the standout feature : at $2 per million input tokens and $10 per million output tokens, it runs roughly half the cost of Opus and about a fifth the cost of GPT-6 Astra.
- A side-by-side rocket launch scene demo showed Kimi finishing in about 6 minutes for roughly $1.95, versus Opus taking about 24 minutes for roughly $10.35, with comparable but not identical visual quality.
- Game and web generation results are mixed , producing playable but visually rough output (a Fall Guys style browser game, a website redesign) that feels competent rather than polished.
Seven tools to build an app. Or just Remy. #
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
How good is Kimi K3.1 Pro at 3D modeling? #
The clearest success case circulating online is a 3D Rubik’s Cube simulation built entirely by the model. It rendered a realistic cube with visible light reflection off each face, true 3D rotation controls, keyboard input support, a scramble function, a timer, and an actual solving algorithm rather than just reversing its own moves. That last detail matters: a lot of models can scramble a cube but fail to genuinely solve it from an arbitrary state. Kimi K3.1 Pro pre-calculated the solution and played it back turn by turn, which is the kind of task that has historically separated closed frontier models from open weights ones.
Moonshot has publicly emphasized this strength, saying the model is excels at building 3D models in Blender and CAD software, and pointing to its performance in 3D rendering and animation more broadly. That claim lines up with a Moonshot-run demo comparing Kimi against Claude Opus 5.5 on a 3D rocket launch scene: both models produced a rocket, launch pad, tower, and smoke effects, with Opus’s version looking slightly cleaner in places, but Kimi finishing in a fraction of the time and cost.
Where does it fall short on complex 3D and spatial tasks? #
Not every demo holds up. One tester asked Kimi K3.1 Pro to model a rope tied into a bowline knot inside Blender, a task that requires web search to understand the knot’s structure, computer control to operate Blender, and genuine spatial reasoning to represent how the rope segments interact. The result got the general shape roughly right (it resembled a figure-8 pattern) but the rope wasn’t actually knotted and wasn’t properly attached to the climbing harness in the scene. A human figure included in the render also looked off.
A follow-up test asked the model to instead build a motion graphic explaining how to tie the same knot, a slightly easier 2D animation task. The first attempt, built in about five and a half minutes, produced a 28-second graphic that gestured at the knot-tying steps but didn’t clearly show where one part of the rope crossed over or under another, meaning it couldn’t actually teach someone to tie the knot. A side-by-side comparison with GPT-6.1 Soul on the identical prompt showed the competing model producing a clearer, more legible result where the over-under rope path was visible and followable.
Asked to simply “make it prettier,” Kimi’s second pass added a drop shadow and smoother camera movement but still didn’t fix the core legibility problem. This pattern, strong surface-level 3D rendering but weaker performance on tasks requiring precise spatial logic, shows up elsewhere too. A WebGL render of a smartphone design looked polished in terms of lighting and material reflections but got the actual product details wrong, including the logo and screen layout, even after the model searched the web for reference images.
Can Kimi K3.1 Pro build games in Unreal Engine and browser engines? #
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Moonshot’s own announcement highlights full project coding support for game engines including Unreal, Unity, and Godot. In practice, hands-on testing has focused more on browser-based game builds than full Unreal Engine projects, but the results give a sense of where the model sits. One tester prompted it with a short, deliberately open-ended request to build a browser-based clone of the multiplayer party game Fall Guys, referencing the real game via web search. The build took roughly 48 minutes and produced a playable 3D obstacle course with multiple animated competitors, moving obstacles like sweepers and sliding gates, checkpoints, a timer, and adjustable settings for speed and jump power.
The tester noted the colors looked overly pastel, collision detection wasn’t fully accurate, and there was visible flickering, but called it a decent result “especially for an open weights model.” That qualifier matters: open weights models have historically lagged well behind closed, proprietary systems at this kind of generative creative coding, so a playable, feature-complete game clone in under an hour is a meaningful data point even with rough edges.
Is Kimi K3.1 Pro actually as good as the benchmarks suggest? #
This is where skepticism has concentrated. Moonshot’s own benchmark charts show Kimi K3.1 Pro beating Claude Opus 5.5 and GPT-6 Astra on tests including Humanity’s Last Exam (70% versus 67.7% and 57.2% respectively), a long-task software engineering benchmark called SWE Marathon, and several internal finance and document-handling tests. Moonshot also reported a sharp drop in hallucination rates compared to the earlier K3 model.
The catch is that most of these numbers come from tests Moonshot ran itself, and some benchmarks were designed internally by Moonshot rather than by a neutral third party. That’s standard practice across the industry, but it means the results deserve a discount. Two independent sources offer a reality check. Artificial Analysis, which runs its own cross-model intelligence index, scored Kimi K3.1 Pro at 55, placing it fourth overall behind Claude Opus 5.5 (58) and just behind Claude Sonnet 5.5 and Meta’s Muse Wonderful (both 56). Arena’s web development leaderboard, based on real user votes between anonymized model outputs, ranked it eighth, behind several Claude and GPT-6 variants but a sharp jump up from where the previous Kimi K3 sat.
Online reaction has reflected that gap between marketing and lived experience. Some users have reported underwhelming results on creative, open-ended tasks like Minecraft builds, even when benchmark scores look strong. The general consensus forming is that Kimi K3.1 Pro is a genuinely capable model, not a fake, but its real-world performance sits a tier below the very best closed models rather than ahead of them.
Why does the price matter more than the benchmark gap? #
Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output tokens through Moonshot’s API, with a faster “High Speed” variant at $4 and $20. That makes it roughly half the price of Claude Opus 5.5 and roughly a fifth the price of GPT-6 Astra, based on published API pricing for those models. In the rocket launch scene demo, that price difference translated into a 3D scene produced for about $1.95 in roughly 6 minutes, versus roughly $10.35 and 24 minutes for Opus.
For builders working on 3D content, prototyping, or coding tasks where “good enough” output at speed matters more than squeezing out the last few points of quality, that cost and time difference is often the more decisive factor than winning a benchmark leaderboard.
Frequently Asked Questions #
What company makes Kimi K3.1 Pro?
One coffee. One working app. #
You bring the idea. Remy manages the project.
Kimi K3.1 Pro is made by Moonshot AI, a Chinese AI lab that previously released Kimi K2 and Kimi K3, and is positioned alongside labs like DeepSeek as a major player in China’s open weights AI scene.
Is Kimi K3.1 Pro open source?
Moonshot has said it plans to release Kimi K3.1 Pro as open weights, meaning the model’s parameters would be downloadable, following the pattern of its earlier Kimi K2 and K3 releases.
How does Kimi K3.1 Pro compare to Claude Opus 5.5 and GPT-6 Astra?
On Moonshot’s own benchmarks it claims to match or beat both models on several tests. Independent rankings from Artificial Analysis and Arena place it close behind the top closed models, generally in the top four to eight range rather than first.
What is Kimi K3.1 Pro actually good at in 3D work?
It has shown strong results on physics-based 3D rendering, demonstrated through an accurate, fully interactive 3D Rubik’s Cube simulation, and in Moonshot’s own side-by-side demo of a 3D rocket launch scene produced much faster and cheaper than a comparable Opus build.
Where does Kimi K3.1 Pro struggle?
Tasks requiring precise spatial logic, like correctly modeling a tied knot in Blender or reproducing exact real-world product details in a 3D render, have shown clear gaps between the model’s visual polish and its underlying accuracy.