license: apache-2.0 base_model:
- empero-ai/Qwen3.8-35B-A3B-Distill
- IsValorum/Qwen3.8-35B-A3B-Distill-MTP-APEX-I-MiniPlus-V2.1-Abliterated-GGUF tags:
- gguf
- blender
- bpy
- code-generation
- 3d
- lora-merged
- abliterated
Qwen3.8-35B-A3B APEX Abliterated + Blender v1 (merged GGUF)
A Blender 5.2 scripting specialist built on Qwen3.8-35B-A3B. A LoRA trained for Blender bpy scripting was merged into the
APEX-I-MiniPlus-V2.1 abliterated GGUF, so it runs in llama.cpp like the original file, with no LoRA flag.
Results (blender-bench, 100 held-out tasks, up to 3 repair rounds with Blender error feedback)
| Base model | This model | |
|---|---|---|
| Final pass rate, thinking off | 5% | 59% |
| Final pass rate, thinking on | 21% | 73% |
| Mean check score, thinking on | 0.29 | 0.85 |
Each task asks for an object (table, chair, office chair, bookshelf, desk lamp, staircase, mug, house, fence, desk setup) with exact sizes. The script runs in headless Blender 5.2 and is checked objectively: collection linked, part counts, dimensions within tolerance, grounded on z = 0, parts resting on each other, no overlaps, materials and modifiers.
Training
- LoRA r=32, alpha=64 on the attention and linear-attention projections, bf16, 4.56M tokens, one pass.
- Data: Blender scripts verified by the checker (solutions from a Qwen3.8-27B model and procedural reference scripts), Blender 5.0 scripts from 3DCode that were re-verified in Blender 5.2 and re-captioned from their renders, plus about 25% general replay data to keep general skills.
- Merged into the GGUF with
llama-export-lora(llama.cpp e85e15c). Merged layers keep the base quant types.
Usage
llama-server -m a3b-apex-abl-blender-v1-merged.gguf -ngl 99 -fa on -c 65536 --jinja
Prompts work best when they state Blender 5.2, Z-up, metres and the target collection name. Give the model the Blender error when a script fails; it fixes most errors within one or two rounds.
Limitations
- Still weak on multi-object scenes (desk setup, house with openings) and occasionally invents
bpy/bmeshAPIs. - This is an abliterated (uncensored) model; apply your own safety layer where needed.
Attribution and licenses
- Base: empero-ai/Qwen3.8-35B-A3B-Distill (Apache-2.0); quantization: IsValorum APEX-I-MiniPlus-V2.1 abliterated GGUF (Apache-2.0).
- 3DCode / 3DCodeBench data (MIT), built on Infinigen: arXiv:2606.01057, https://github.com/gaoypeng/3dcodebench.
- Data generation helper model: DavidAU Qwen3.8-27B TURBO Cold-Fusion NEO-CODER-MAX (Apache-2.0).