gemma-4-12b-constraintkit

gemma-4-12b-constraintkit

A 12B Gemma fine-tune that turns plain English into a parts-and-mates program for the constraint-kit CAD system, and says so when a request is out of scope, for makers and engineers.

Made by shib79 on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: gemma-4-12b-constraintkit is a Gemma 4 12B fine-tune by shib79 for constraint-kit, a separate open-source text-to-CAD system. You describe a part or assembly in plain English. The model reasons first, then writes a small program that lists parts, their parameters and the mates between them. constraint-kit then builds exact solids, solves the assembly and exports files such as STEP. If a request is outside what the system models, the model answers with an unsupported note and a reason, instead of guessing.

Who it suits: It suits mechanical engineers, makers and developers who want CAD output that can be checked against real geometry, and researchers who study training with a verifier.

Title: Who uses this model
Industry | Who uses it | How they use it | Example
Engineering | Mechanical engineers | Draft parts and assemblies from a text description | Ask for three spacers stacked on top of each other
Making | Hobbyists | Turn an idea into a CAD program | Describe a part, then build it with constraint-kit
AI research | Researchers | Study reward signals from a CAD verifier | Compare the supervised stage with the GRPO stage

If you want a mesh or a finished drawing straight from the model, you can skip it, since it writes a program that constraint-kit must compile.

Where to use it:

• llama.cpp, through llama-server, using the GGUF file
• constraint-kit, the separate project that compiles the model's program
• Windows, macOS or Linux with Python, for downloading the files

Title: Jobs it does
Use case | Who | What they get back
Text to CAD program | Engineers | A program of parts, parameters and mates
Honest declines | Teams that need trust | An unsupported answer with a reason for requests out of scope
Visible reasoning | Reviewers | A thinking trace before each program
Export through constraint-kit | Makers | STEP or GLB files, a parts list, mass and drawings

How to set it up: You need Python for the download, a llama.cpp build and room for the 9.8 GB Q6_K GGUF file. The card also offers a 251 MB LoRA adapter on the unsloth Gemma 4 12B base. The repository includes train_sys.txt, the system prompt that training used. The card marks it as required, since training never used a different prompt.

• Python with pip
• A llama.cpp build that includes llama-server
• About 9.8 GB of free disk space for the Q6_K file
• constraint-kit, to turn the program into real solids

1. Install huggingface_hub.
2. Download the repository files, including train_sys.txt.
3. Start llama-server with the GGUF file.
4. Send the training system prompt with your plain-English request.

pip install -U huggingface_hub
hf download shib79/gemma-4-12b-constraintkit-GGUF

llama-server -m gemma4-12b-constraintkit-grpo-Q6_K.gguf -c 8192 -ngl 99 --port 8086

How to use it: Once the server runs, send it the system prompt from train_sys.txt and then your request. The model replies with its reasoning and a program, which you pass to constraint-kit. The visible part of the card ends after the server command, so check the full card for the exact call.

Title: Main command pieces
Tool | What it does
hf download | Fetches the files from the repository
llama-server -m | Starts the server with the model file
-c 8192 | Sets the context size
-ngl 99 | Puts the model layers on the GPU
--port 8086 | Sets the port the server uses

Example: You ask for three stacked spacers, each 12 mm across with a 5 mm bore and 8 mm tall. The model writes a program with three spacer parts and mates joining each top to the next bottom. If you ask for a spiral bevel gear, it declines and says only straight bevel gears are modeled.

For newcomers and experts: CAD means computer-aided design, the software that draws exact parts. A first small test is the stacked spacers request, since the card uses that kind of simple example.

Training had two stages, supervised tuning and then GRPO, which used the CAD verifier as the reward, with no reward model and no language-model judge. The card scores a build by comparing its volume and bounding box with the held-out target, with a 2% tolerance. The card says GRPO halved silent wrong-variant builds, from 2 of 10 to 1 of 10, and raised correct declines to 9 of 10.

Plain English request
Model reasons, then writes a program
constraint-kit solves the mates
STEP or GLB files, parts list, mass, drawings

Good to know: The author reports 93% of builds verified against geometry and 87% honest declines on held-out probes at the supervised stage. On fresh requests it reports 57% honest builds, against 16% for a prompted frontier model. Training cost about $35 in total. These are the author's own tests on small sets, so treat them as signals to check. The card's gallery pictures came from constraint-kit's hosted planner, not from this model.

Title: Honest builds on fresh requests, in percent
This model: 57
Prompted frontier model: 16

• License: Apache 2.0.
• Cost to run: none named, apart from the hardware.
• Data it reaches: the files come from Hugging Face, and the model then runs on your own computer.
• Not for: parts outside its scope, such as spiral bevel gears, which it should decline.

On Hugging Face, 13 November 2026

• Task: text to 3d
• Library: gguf
• Parameters: 11.9B
• Weights: about 6 GB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 524
• Likes: 9
• License: apache-2.0
• Languages: en
• Based on: google/gemma-4-12B-it
• Tags: gguf, safetensors, text-to-cad, parametric-cad, b-rep, step, 3d, mechanical-engineering, engineering, gears, honest-refusal, grpo

Links

• Model: https://huggingface.co/shib79/gemma-4-12b-constraintkit-GGUF
• Files: https://huggingface.co/shib79/gemma-4-12b-constraintkit-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download shib79/gemma-4-12b-constraintkit-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.