For Momokji Core owners
Recommended AI models.
If you run your own Momokji Core with a local Ollama server, this is the model we suggest. Every app on your Core shares it, so we lead with one overall pick and then note where individual apps need something more.
A Momokji Core uses one language model for all of its apps: the Core's inference model, chosen in Momokji Manager → AI → Chosen. These recommendations are revisited over time as new models are released and as we measure more hardware. Each entry shows when it was last recommended and what it was measured on.
Overall pick
gemma4:12b
Answers in about 0.4 seconds once loaded, at about 100 tokens per second, and gave the best companion answers of the models we tested.
- Size
- About 7.6 GB
- Quantization
- Q4_K_M
- Tools & vision
- Tools, Vision
- Reply once loaded
- About 0.4 s
- Speed once loaded
- About 100 tokens/s
- Cold load
- Not yet measured
Last recommended: · Measured on NVIDIA RTX 4090 · Ollama 0.40.6 · November 2, 2026
Smaller alternative
gemma4:e4b
Much faster once loaded, at about 220 tokens per second, but its companion answers were noticeably weaker.
- Size
- About 6.6 GB
- Quantization
- Not yet measured
- Tools & vision
- Not yet confirmed
- Reply once loaded
- Not yet measured
- Speed once loaded
- About 220 tokens/s
- Cold load
- Not yet measured
Last recommended: · Measured on NVIDIA RTX 4090 · Ollama 0.40.6 · November 2, 2026
Heavier option
qwen3.8:27b
Good answers, but slow to load: about 22 seconds from cold before the first reply.
- Size
- About 17.7 GB
- Quantization
- Not yet measured
- Tools & vision
- Not yet confirmed
- Reply once loaded
- Not yet measured
- Speed once loaded
- Not yet measured
- Cold load
- About 22 s
Last recommended: · Measured on NVIDIA RTX 4090 · Ollama 0.40.6 · November 2, 2026
Keeping companions quick
- Loading is what feels slow. What makes a companion feel slow is the model loading, which took 8 to 22 seconds in our tests, not the model's size once it is running.
- Keep one model loaded. Use Ollama's keep-alive so the chosen model stays in memory between conversations.
- Don't make big models compete. Avoid several large models taking turns on one GPU. Each swap means another load, and every app on the Core waits for it.
Notes by app
Couch PotAIto · Spuddy
Needs: Short, witty replies on the TV. gemma4:12b keeps replies quick once loaded. Fast, short answers matter more here than long reasoning.
Last recommended:
ReelVision · ReelVision companions
Needs: Conversation and tool use; image understanding helps. gemma4:12b supports both tools and vision, so one model covers chat, tools and images.
Last recommended:
Who DatPlanned
Needs: Requires vision. Needs a vision-capable model. gemma4:12b is vision-capable; Who Dat itself has not yet been measured.
Last recommended:
Recommendations version 2026-11-02. Measurements use a short companion-style prompt; your hardware will give different numbers. Read more about AI & AnaMind.