Take a hand-drawn sketch, photograph it, and send it to Kimi. A few minutes later you have a downloadable 3D model file, and at no point in the process did you open a modeling application. On October 24, Moonshot AI released K3.1 Pro and, alongside it, quietly switched on a feature called 3D Make on kimi.com. The model does not merely imagine what an object looks like. It opens Blender, writes a script, runs that script in the cloud, and hands you the finished geometry. The wall that has stood between ordinary creators and 3D modeling for decades is being dismantled in a way very few people predicted: not through a smarter brush or a better text-to-image model, but through an agent that writes code.
How 3D Make Actually Works: Writing Scripts, Not Meshes
To understand what makes 3D Make different, start with what it does not do. It does not generate a mesh directly, and it does not run the familiar diffusion-style pipeline that turns a prompt into a blocky, over-textured blob. According to what the kimi.com interface shows, you provide a text description, a sketch, or a reference image. The model infers the object's shape, proportions, and spatial relationships, then autonomously writes executable scripts for professional modeling software such as Blender, FreeCAD, and SolidWorks. Those scripts are uploaded to the cloud, executed there, and the results are returned to you. In effect, an AI is remotely operating professional modeling software, producing geometry through parametric construction rather than statistical guessing.
The value of this route lives in the word "executable." What comes back is not a rendering but a genuine modeling artifact: dimensions carry numeric values, geometry has constraints, and everything can be refined afterward. Moonshot AI's official blog states that 3D Make requires no MCP connectors and no external tool authorization; the model writes and runs the modeling scripts directly, with the entire pipeline closed inside the product. The interaction stays deliberately plain. You can revise through conversational instructions, attach an annotation to a specific element for targeted adjustments, or edit manually. Say "make this cylinder two centimeters taller and set the wall thickness to three millimeters," and the model recomputes and regenerates. You keep working at the level of design intent rather than patching a pile of triangles by hand.
There is a quieter benefit that deserves mention: determinism. Run the same script twice and you get the same result, which is a fundamentally different experience from generative models where every roll of the dice produces a new variation. For anything involving deliverables, reproducibility is not a luxury. It is productivity.
Asking a model to write CAD scripts is not a new idea, and the hard part was never the idea. The difficulty sits in a translation task that fails silently: whether that wobbly line in a sketch is a fillet or a chamfer, whether two edges are linked by a constraint or remain independent dimensions. Get it wrong and nothing errors out; you simply receive a model with the wrong shape and have to notice on your own. What 3D Make changes is where that hardest judgment step happens, moving it into a model that was specifically post-trained for hands-on capability rather than conversation.
What Powers It: The K3.1 Pro Post-Training Ledger
A script-driven product only works if the model behind it can actually be trusted to write and run scripts, and that is what the same-day release of K3.1 Pro provides. By the official account, it is a mixture-of-experts model with 2.8 trillion total parameters and 104B parameters activated per inference, with a context window of 1 million token capacity. Moonshot calls it the strongest 3T-class open-weight model to date.
The ledger is worth reading line by line. The official blog lists coding, 3D creation, financial research, and agentic capabilities as the model's designated strengths, and the benchmark most relevant to how 3D Make operates is Terminal-Bench 4.0, where K3.1 Pro scores 65.7%. That test measures how well an agent performs real operations inside a terminal environment: write a script, run a tool, read the output, and iterate. That loop is almost exactly the loop 3D Make runs on every request. The 65.7% score puts it ahead of GPT-6 Astra at 58.2% and within 0.7 percentage points of Claude Opus 5.5. On the Artificial Analysis intelligence index it earns 55 points, ranking fourth among 215 models and first among open-weight ones. Third-party leaderboards invite legitimate debate about methodology, but the pattern is hard to miss: this post-training round was built, first and foremost, to make the model good at doing things rather than merely talking about them.
Another number matters even more for this particular use case. Official stress tests report that the model's claim-level hallucination rate dropped from 23.4% in K3 to 8.3%. In a script-driven 3D workflow, that improvement is not a nice-to-have. The model is accountable for every line it writes, and a single fabricated API call is enough to break an entire script and frustrate an entire afternoon. Cutting hallucinations by roughly two-thirds is what moves the script route from demo videos into something a working professional can lean on.
The 1 million token context is not decoration in this pipeline either. For ordinary users, it means a full reference document, the entire revision history of a design, or a whole asset inventory can sit inside the conversation while the model works, without the model guessing your intent from scratch each time. For advanced users, feeding in the modeling scripts of an existing project and asking the model to extend it with new parts is precisely the scenario this window was sized for.
Pricing, finally, is on the table. The standard API costs 2 dollars per million token of input and 10 dollars per million token of output. The Highspeed tier costs twice as much and delivers output two to three times faster. For developers planning to batch-run modeling scripts, that is at least a cost structure you can compute against before committing, rather than a marketing promise of "dirt cheap" that falls apart on the first invoice.
From Sketch to Printable: Workflow and Export Formats
String the workflow together: provide a description or a sketch, the model infers geometry and proportions, writes the script, runs it in the cloud, and produces a file. The output can be downloaded or previewed online, and multiple export formats are supported. GLB, glTF, and OBJ are all available, and the official resource page also mentions FBX, STL, USDZ, and BLEND depending on your workflow. Moonshot's own positioning is blunt: generate a 3D-printable model from a single sentence.
The phrase "3D-printable" deserves its own paragraph. STL remains the universal input format for consumer printers, and whether a model can actually be printed depends on whether it is a watertight solid with sensible wall thickness. Models produced through parametric scripts have a structural advantage here: the geometry is generated under numeric constraints with clean boundaries, rather than grown out of a voxel field, which is where broken surfaces and non-manifold edges typically come from. That said, run any AI-generated model through a slicer before sending it to the printer. Overlapping shells and unsupported overhangs can still slip through, and there is no reason to mythologize the pipeline just because the script was written by a model.
For editing, the product offers three paths: conversational instructions, annotations attached to individual elements, and manual editing. After a dimension or geometry change, the model recomputes and regenerates, which means your edits persist at the level of design intent. A parameter you changed survives the next regeneration instead of being flattened into a manual fix you have to redo. The official materials also mention scene assembly and game asset production. The former lets you combine multiple generated objects into one scene, useful for stages, rooms, and level layouts. The latter is a concrete workload reduction for indie developers who need crates, props, and furniture by the dozen and have no business hand-modeling each one.
One small detail matters more than it sounds: inline preview. Every revision can be spun around and inspected in the browser immediately, with no export-import detour through other software, which compresses the feedback loop to seconds. How fast a creation tool feels is largely decided by the speed of that loop.
Why This Is Not Text-to-3D
The past two years produced a wave of text-to-3D tools, and AI 3D generation briefly became a startup gold rush. That route uses diffusion models or generative networks to output meshes or voxels directly. It is fast, and speed is genuinely its virtue, but the artifacts are usually models that merely look right: chaotic topology, redundant triangles, no concept of real-world dimensions, and a long distance from engineering usability. It hands you an attractive concept image. It does not hand you a production file.
General-purpose assistants had already drifted toward CAD before this. ChatGPT and its peers were asked to write OpenSCAD scripts or drive CAD kernels directly, but the industry's recurring complaint was "what you write is not what you get": the script runs, yet the output does not match the description, and tweaking a single sentence reshapes everything beyond recognition. The difference in 3D Make's approach is that it did not stop at having a model write scripts. It connected large-model capability to professional DCC toolchains such as Blender, FreeCAD, and SolidWorks, so scripts run on mature modeling kernels with real feedback that can be read back and acted on. Observers have called it one of the rare instances of genuine engineering follow-through by a Chinese AI lab in 3D content production, a field where announcements have often outrun implementations.
The two routes are not fighting to the death. Text-to-3D suits concept drafts and stylized assets where speed matters more than precision. Script-driven modeling suits anything that needs dimensions, constraints, editability, and printability. What 3D Make adds is a bridge between "impressive demo" and "production tool," a plank that has been missing from the industry floor for years. How much weight that plank can carry is a question only subsequent iterations will answer.
Open Weights by November 4: What Developers Should Watch
For developers, the date worth circling is not October 24 but November 4. Moonshot AI's official blog explicitly commits to releasing the full model weights by November 4. As of this writing on November 3, the weights have not appeared. What that means in practice is that everything you can use today is the productized entry on kimi.com, and within days a model with 2.8 trillion total parameters and 104B activated parameters could be running on your own inference cluster.
Two pieces of reality belong next to that promise. First, a mixture-of-experts model of this scale is not light on memory or scheduling, and local full-precision deployment is out of reach for individual enthusiasts. Quantized releases and distributed serving setups will be where the community concentrates its effort next, and the waiting period is exactly the right time to run the hardware math. Second, Moonshot proactively disclosed a safety decision: the open-weight version removes cybersecurity data that had been mixed into the general corpus, and attack-related capabilities drop by roughly 9.5 percentage points as a result. The open weights and the hosted service are therefore not identical at the capability boundary, and teams planning security research or red-team evaluations should know that before they start, or their findings will not transfer to the hosted product.
Teams planning to bet on this route can also prepare something in advance: the ceiling of a script-driven product depends on how deeply the model understands the modeling ecosystem. Once the weights are out, domain fine-tuning gets a real handle. Mold design, sheet metal parts, and architectural components are all terminology-heavy niches that can be fed with proprietary corpora. A vertical the product does not cover today is not necessarily a permanent blank.
The launch week was not all applause either. According to a hands-on report from NetEase Technology, on the third day after release Kimi publicly apologized for rate limiting and abnormal quota consumption, compensating members with an additional 25% of monthly quota. The lesson is unglamorous: productized capability and reliable service at scale are separated by an operational gap, and no benchmark score substitutes for stable supply. The direction is worth waiting for, the weight commitment is worth tracking, and patience costs nothing. Let the weights land on November 4, let the community run its tests, and then decide whether 3D Make belongs in your stack. Rushing the verdict helps no one.
Related reading