entity12208/ editorai:v2

327 17 hours ago

The official Ollama models for the Editor AI mod in Geometry Dash!

tools 
ollama run entity12208/editorai:v2

Details

4 days ago

482cbe1f1630 · 986MB ·

model
qwen2
·
1.54B
·
Q4_K_M
system
You are EditorAI's level-design model for Geometry Dash. You have tools available. USE THEM PROACTIV
params
{ "num_ctx": 8192, "stop": [ "<|im_end|>", "<|endoftext|>" ], "tempe
template
{{- if or .System .Tools }}<|im_start|>system {{- if .System }} {{ .System }} {{- end }} {{- if .Too

Readme

EditorAI on Ollama — model family

AI models for the EditorAI Geode mod that generate Geometry Dash levels. All models live under the entity12208/editorai namespace on the Ollama registry.

Which one should I run?

Tag Size (Q4_K_M) Base model Fine-tune Best for
entity12208/editorai:v4-14b 8.4 GB Qwen2.5-14B-Instruct QLoRA 4-bit, EAS-native Flagship. Best quality. Needs ≥12 GB VRAM.
entity12208/editorai:v3-7b 4.4 GB Qwen2.5-7B-Instruct QLoRA, EAS-prompted Best balance. Runs on 6–8 GB VRAM (RTX 3050+).
entity12208/editorai:v2 941 MB Qwen2.5-1.5B-Instruct QLoRA, JSON-trained Lightweight fallback. Runs anywhere. Tool-use works w/ mod fallback parser.
entity12208/editorai:experimental 941 MB Qwen2.5-1.5B-Instruct Earlier V2 candidate Kept for backward compat. Use :v2 instead.
entity12208/editorai:mini 398 MB Qwen2.5-0.5B-Instruct Early experimental Archive only. Not recommended.

Default pick: v3-7b for most users · v4-14b if you have ≥12 GB VRAM and want the best results · v2 if you have GB VRAM or want something tiny.

Install

# Pull whichever tag fits your hardware:
ollama pull entity12208/editorai:v4-14b
ollama pull entity12208/editorai:v3-7b
ollama pull entity12208/editorai:v2

In-mod: AI Settings → provider = ollama, model = (the tag you pulled), Platinum = off.

Speed (Q4_K_M, generation t/s)

GPU v2 (1.5B) v3 (7B) v4 (14B)
RTX 3050 6 GB Mobile 50 25–35 partial CPU offload, ~3 t/s
RTX 3060 12 GB 100+ 60–80 30–40
RTX 4060 Ti 16 GB 150+ 70–90 40–55
RTX 3090 / 4090 250+ 80–120 50–90
Apple M3 Pro (Metal) 80+ 30–45 20–30
CPU only 6–8 2–3 0.5–1

Output format

All EditorAI models output to the EAS (EditorAI Script) format — a line-based DSL that the mod parses into actual level objects. The mod also accepts JSON output as a fallback (older models and frontier APIs like Claude/GPT do this). Both formats are tolerant of missing analysis and objects fields (metadata-only updates are valid).

Model Output format Notes
v4-14b EAS-native (trained on EAS verbs) Best EAS quality
v3-7b EAS via in-context examples Solid EAS + JSON fallback
v2 / experimental JSON Mod has fallback parser to handle their quirks
mini JSON Same caveats as v2

Tool use

v2 / v3 / v4 all support the mod’s 7 tools (get_level_length, analyze_level, think, web_search, search_newgrounds, get_newgrounds_song, download_level). They emit tool calls in Qwen2.5’s native <tool_call> format, which the mod parses and routes.

License

All EditorAI models are released under Apache-2.0, inherited from the Qwen2.5-Instruct base. You can use them commercially or non-commercially with attribution.

Source