# Maris3D-SLM-Blender-0.5B v0.1 — Ollama Modelfile (default file) # # Local build: ollama create maris3d-slm-blender-0.5b -f Modelfile # After publish you can also pull straight from Hugging Face — ALWAYS with the :Q8_0 tag: # ollama run hf.co/slm4apps/Maris3D-Blender:Q8_0 # (The untagged name resolves to the unmerged BASE q4_k_m build — NOT this merged default. # An hf.co pull also bypasses this Modelfile — see SERVE.md for the remote Modelfile recipe.) FROM ./maris3d-slm-blender-0.5b-edit-q8_0.gguf # Serving parameters (full recipe in SERVE.md) PARAMETER temperature 0 PARAMETER repeat_penalty 1.0 PARAMETER num_ctx 8192 # Mandatory generation cap. Ollama's default generation is UNLIMITED; prompts outside # the trained tool-loop (plain questions) then repeat forever. 800 covers the largest # observed tool-call bursts. PARAMETER num_predict 800 # Always send the 26-tool `tools` array with each request (see SERVE.md): without it # the model replies with a canned formula line instead of an answer. # Identity line — canary A/B decided 8 Nov 2026: WITH the line pass 18.3% / usable # 65.0% vs 16.7% / 63.4% without (unparsed 15→9, edit bucket 0/10→3 pass). Ships on. SYSTEM You are SLM4Blender-0.5B, a compact local model by slm4apps that builds and edits Blender scenes through tool calls.