Llama: only local infra (e.g., Llama 3.2 3B at 124 tok/s on a jailbroken PS5)
CodeRead the post
x.com · link re-checked 2026-10-16
Gemma 4 31B
By Google. 2 results on 2 boards, copied as published. Different boards, different metrics: don't add them up.
License (as listed)Apache 2.0
Hosting regionNot confirmed
Boards2
Code
No result on the boards we captured
Agentic / terminal
No result on the boards we captured
Websites
- 1368LMArena WebDev Arena
gemma-4-31b · Arena score · board #113
LMArena WebDev (agentic front-end tasks) · as of 2026-10-22 ↗
Design
No result on the boards we captured
Android
- 37.1%Android Bench 1.0 tasks (100 tasks)
Gemma 4 31b IT · Score %
not_stated · as of 2026-08-11 ↗
iOS
No mature public leaderboard
Available in
No plan row in our pricing data names this model version.
Use cases with this model
No sourced use case names this exact model version.
Same family: Gemma 9
The follow-up: v2 accuracy went "0.513 to 0.804 … the student now beats its teacher"
CodeRead the post
x.com · link re-checked 2026-10-16
Fine-tuning a decision model on a gaming rig: "Gemma 4 E2B + LoRA in @UnslothAI Studio. 16 GB of VRAM, about 12 minutes"
CodeRead the post
x.com · link re-checked 2026-10-16
"Gemma-4-12b fine tuned to play snake … about 500ms per decision" on a 70W GPU
CodeRead the post
x.com · link re-checked 2026-10-16
Real-time voice agent "with zero US servers" — Gemma 4 on Scaleway
CodeRead the post
x.com · link re-checked 2026-10-16
Running Gemma 4 offline in the Antigravity SDK via LiteRT: "Zero API costs"
CodeRead the post
x.com · link re-checked 2026-10-16