Maris3D SLM Blender 0.5B (v0.2.1)

A 0.5B model that builds Blender scenes from plain words, offline.

GitHub Release License: Apache-2.0 Blender 4.2+

A snowy street with houses, a bus, a snowman and a Christmas tree, built in Blender by Maris3D

Install in one command

No need to clone anything. This installs the Blender add-on, downloads the two model files below (sha256-checked) and tests them in your Blender. Needs Python 3.9+ and Blender 4.2+.

macOS / Linux Windows (PowerShell)
curl -fsSL https://raw.githubusercontent.com/slm4apps/maris3d/main/install.sh | sh powershell -ExecutionPolicy ByPass -c "irm https://raw.githubusercontent.com/slm4apps/maris3d/main/install.ps1 | iex"

Then: maris3d serve start, open Blender, press N in the 3D View, pick the Maris3D Copilot tab, type a sentence and press Run. Try:

Build one thing Build a whole scene Change what is there
a small house a park with a bench, a tree behind the bench and a lamp next to the bench Make the bench blue. then Move the lamp to the left of the bench.

You describe a scene in a sentence. The model answers in plain words: which things, how many, what they are made of and where they stand. The Maris3D app turns those words into exact Blender tool calls, builds the scene, reads it back and checks it. The model never writes a number; code computes every size and position.

It runs fully offline on your own machine: one 531 MB base file plus a 12 MB adapter, on a 4-core laptop CPU at 28.5 tok/s in 876 MB of memory.

On held-out sets it plans the right things 97.3 % of the time, designs a new thing's parts that pass the app's rules 95.7 %, makes the word edit asked for 90.4 %, and replays the 204 requests of the 0.1 demo scene 188 exactly (92.2 %). Read the limits section before you use it.

Uncensored by training. The fine-tune leaves no refusal behavior: dark game-art requests (weapons, gore, horror props, standard 3D art fare) are planned like anything else. 0 of 18 refused in our probe.

Files

File What it is
maris3d-slm-blender-0.5b-q8_0.gguf The base (531 MB): the 0.1 weights, with Maris3D metadata.
maris3d-slm-blender-0.5b-v0.2.1-lora-f16.gguf The adapter (12 MB): plans, parts, word edits, talk.
Modelfile The Ollama entry maris3d-slm-blender-0.5b, with the settings the app sends.
MANIFEST.json, SHA256SUMS The current file of each role with its sha256, read by maris3d update.

What is new in 0.2

  • Words, not numbers. 0.1 wrote tool calls with coordinates. 0.2 answers in one small word language (things, relations such as behind or on top of, sizes such as half or double), and the app's compiler writes the tool calls.
  • One adapter for everything: planning, designing a new thing's parts, word edits and talk, instead of 0.1's separate edit and planner adapters.
  • Animation by the app's code: "a ball bouncing twice", "a moon going round a planet".
  • Edits by words. "Make the table darker", "move the lamp next to the sofa": the model says what changes, code applies it and checks it.
  • Talk. Ask about the scene and it answers from what is there.
  • Breaking change: 0.2 needs the 0.2 app. 0.1 stays available as Maris3D-Blender v0.1.0 and at tag v0.1.1 on GitHub.
  • 0.2.1: The same weights as 0.2.0. The files now carry the Maris3D name and identity; maris3d update fetches both again (543 MB).

See it build

The finished demo street, rendered in Blender: a slow zoom into the hero house and its fireplace.

the finished street, a slow zoom into the hero house and its fireplace

Screen recordings of the street being built (Blender 5.2, recorded with 0.1, sped up). Each thing is one typed request. The 204 requests of these demo scenes, replayed through the 0.2 app: 188 built exactly (right thing, place and size).

the room, from the first table to the walls the play park, from the first blade of grass
the tree, the snowman, the gift, the bus and the people More recordings and renders of every thing in the scene are in the GitHub repository.

Maris is the add-on's mascot (optional): she weaves while the copilot builds and reacts when it finishes. Here she stands beside groups of things from the demo street in a different outfit, skin, hair and pose each time; every part appears in turn, then a 360 turn.

the tree and the snowman, with Maris the gifts at the door, with Maris
the kitchen, with Maris the park friends, with Maris

Ways to run it

  • The Maris3D app (recommended). maris3d install fetches the base and the adapter; maris3d serve start runs llama-server with them. maris3d update later fetches only the files that changed.
  • llama-server by hand: llama-server -m maris3d-slm-blender-0.5b-q8_0.gguf --lora maris3d-slm-blender-0.5b-v0.2.1-lora-f16.gguf --lora-init-without-apply --jinja -c 8192
  • Ollama: ollama create maris3d-slm-blender-0.5b -f Modelfile (that is :latest, the name the app uses; ollama cp maris3d-slm-blender-0.5b maris3d-slm-blender-0.5b:0.2.1 keeps this version by name), then pick Ollama in the app's settings.

Serve with temperature 0, a context of 8,192 and repeat penalty 1.05 (the app sends these on every request).

Results

Held-out sets

Sets the model never trained on, scored by outcome through the app's own code.

What Result
Plan from one sentence: the things named are right 97.3 % (3,033 things in 716 held-out sentences)
the relation (behind, next to, on top of ...) is right 97.0 %
the thing it is placed by is right 95.8 %
Parts of a new thing: the answer parses 99.9 % (1,268 held-out names)
the design passes the app's rules 95.7 %
Word edits: the change asked for is the change made 90.4 % (726 held-out change requests)
Talk: the reply is right 95.8 % (118 held-out talk turns)
Talk: things invented that are not in the scene 0 in 383 turns
The 0.1 demo scene, its 204 requests replayed through the 0.2 app: built exactly (right thing, place and size) 92.2 % (188 of 204)

0.2 on the held-out sets

The 180-request benchmark

Our internal suite of real requests across single objects, multi-part scenes, arrays and rings, parenting, geometry nodes, animation and edits (the public benchmark is being prepared), with the same frozen scorer for every model. pass % = fully correct; usable % = correct or correct with small fixes. Every row is n=180: 150 build requests and 30 edit follow-ups. The other models' rows are the ones measured for 0.1; the API and CLI arms ran on the providers' machines, the CLI arms inside their own agent scaffolding, so read the table as tiers.

Model ran on build pass / usable (n) edit pass / usable (n)
claude-haiku-5.5 (CLI) remote API 52.7 / 89.3 (150) 70.0 / 96.7 (30)
gpt-6-luna (CLI) remote API 46.7 / 78.0 (150) 76.7 / 90.0 (30)
qwen3.8-27b (API) remote API 31.3 / 64.0 (150) 73.3 / 93.3 (30)
Maris3D 0.2 4-core laptop CPU 24.7 / 80.7 (150) 16.7 / 46.7 (30)
Maris3D 0.1 4-core CPU VM 24.0 / 73.3 (150) 43.3 / 86.7 (30)
deepseek-v4.1-flash (API) remote API 12.0 / 30.0 (150) 43.3 / 66.7 (30)
qwen3.5-2b 4-core CPU VM 10.0 / 50.0 (150) 23.3 / 46.7 (30)
qwen3.5-0.8b laptop CPU 4.7 / 50.7 (150) 6.7 / 43.3 (30)
qwen3-0.6b 4-core CPU VM 2.0 / 14.7 (150) 6.7 / 36.7 (30)
qwen3-1.7b 4-core CPU VM 1.3 / 6.0 (150) 23.3 / 46.7 (30)

Overall, 0.2 is 23.3 % pass / 75.0 % usable. The 0.1 rows are the 0.1 card's run (27.2 / 75.6); a later 0.1 run with the router picking the adapter per request gave 28.3 / 88.9 (BENCHMARKS). On the 150 build requests 0.2 gives the most usable scenes of the local models (80.7 %, 0.1 73.3 %). Untrained small models (plain Qwen2.5-Coder-0.5B, the 1.5B and 7B coders, Hammer2.1-0.5b, gemma-3-1b-it) score 0.0 on this benchmark.

Maris3D alongside the models we tested

By mode, pass % / usable %; edit rows are follow-ups in a session:

Model single multipart arrays/rings parenting geonodes animation edit
claude-haiku-5.5 (CLI) 52.5/97.5 30.0/80.0 90.0/100.0 55.6/72.2 91.7/100.0 33.3/86.7 70.0/96.7
gpt-6-luna (CLI) 37.5/75.0 46.7/90.0 70.0/80.0 66.7/77.8 50.0/66.7 30.0/73.3 76.7/90.0
qwen3.8-27b (API) 22.5/65.0 6.7/36.7 45.0/65.0 50.0/77.8 75.0/91.7 30.0/70.0 73.3/93.3
Maris3D 0.2 22.5/97.5 10.0/83.3 40.0/70.0 5.6/66.7 0.0/50.0 53.3/83.3 16.7/46.7
Maris3D 0.1 10.0/70.0 13.3/76.7 40.0/60.0 22.2/77.8 83.3/83.3 20.0/76.7 43.3/86.7
deepseek-v4.1-flash (API) 7.5/30.0 0.0/13.3 15.0/20.0 27.8/38.9 33.3/66.7 10.0/33.3 43.3/66.7
qwen3.5-2b 2.5/40.0 6.7/46.7 25.0/25.0 27.8/72.2 16.7/25.0 0.0/80.0 23.3/46.7
qwen3.5-0.8b 5.0/60.0 0.0/60.0 15.0/40.0 11.1/66.7 0.0/16.7 0.0/40.0 6.7/43.3
qwen3-0.6b 5.0/42.5 0.0/0.0 0.0/0.0 5.6/11.1 0.0/8.3 0.0/6.7 6.7/36.7
qwen3-1.7b 2.5/10.0 0.0/0.0 0.0/5.0 0.0/0.0 0.0/0.0 3.3/13.3 23.3/46.7

0.1 and 0.2 by mode

Public benchmarks (measured on 0.1)

0.1 ran 3DCodeBench (212 of 212 scenes) and BFCL v4 simple, python (49.25 %). Both score a model that writes tool calls itself; in 0.2 the app's compiler writes the calls, so they were not run again.

public benchmarks, 0.1

Speed

Base + adapter in llama-server, as the app serves it, on CPU only:

ran on generation prompt reading memory
4-core Intel laptop CPU (i7-6700HQ), 4 threads 28.5 tok/s 142 tok/s 876 MB

Measured with a 555-token prompt and 128 generated tokens, three runs (28.2 to 28.8 tok/s). One benchmark request, from the sentence to a checked scene in Blender, takes 16.0 s (median of 180) on that CPU.

The base file is the one 0.1 shipped. Measured for 0.1, the base alone runs at 96.7 tok/s on an Apple-silicon Mac (Metal), 141.5 tok/s on a laptop with a 4 GB GTX 980M, and 62.1 tok/s on a 4-core CPU VM:

the base on three machines, 0.1

speed across the models tested for 0.1

Training

Base: Qwen2.5-Coder-0.5B-Instruct (Apache-2.0). The adapter is a LoRA trained on text-only conversations in the word language.

Limits

  • No lights, cameras or text from a plain request yet; an animation that names a camera or a light ("a camera circling the cube") gets a real one from the app's animation code.
  • It writes the word language; the Maris3D app's compiler turns that into Blender tool calls.
  • Edit follow-ups on the 180-request benchmark: 46.7 % usable (0.1: 86.7 %).
  • Where it is weaker, mode by mode:

where 0.2 fails, by mode

Support the project

Maris3D is built and trained by one person on home machines. If it is useful to you, you can support the next versions on Ko-fi or Buy Me a Coffee.

Licence

Apache-2.0 for the weights and code (see LICENSE and NOTICE). The Blender add-on, which runs inside Blender, is GPL-3.0-or-later. Base model Qwen2.5-Coder-0.5B-Instruct is Apache-2.0 (Alibaba Qwen team). This project is not affiliated with, endorsed by, or supported by the Blender Foundation. "Blender" is a trademark of the Blender Foundation.

Downloads last month
1,162
GGUF
Model size
0.5B params
Architecture
qwen2
Hardware compatibility
Log In to add your hardware

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for slm4apps/Maris3D-Blender

Finetuned
(119)
this model

Collection including slm4apps/Maris3D-Blender

Evaluation results

  • full-scene pass % on Maris3D 180-request benchmark
    self-reported
    23.300
  • usable scene % on Maris3D 180-request benchmark
    self-reported
    75.000