§ ios · testflight beta
Unbound for iOS.
A more stable runtime, working vision support, and faster generation than the browser build.
0
servers · runs on your device
0
refusals · no policy layer
504.5 MB
default model · cached in OPFS
§ chat
Chat, uncensored.
Ask anything and get a straight answer. The model runs on your device, so your conversations never leave it.
Open Chat§ image generation
Generate images, locally.
Describe what you want and get an image back in seconds, made right on your device. Nothing leaves your machine.
Open Image Studio100% privateNo accountWorks offline
First run downloads the model once, then everything works offline.
§ benchmarks
Benchmarked against the base model.
vs base
gemma-4-e4b-itRefusal rate
AdvBench 520 · LLM judge
base
↓
unbound
Useful-compliance
base
↑
unbound
Hallucination
on harmful prompts
Hallucination
on harmful prompts
↑0.00%0.00%
+11.73 ptsCoherence
benign prompts
Coherence
benign prompts
·0.000.00
0TruthfulQA mc2
--limit 100
TruthfulQA mc2
--limit 100
↑0.0000.000
+4.7 ptMMLU
--limit 100 · 61 subtasks
MMLU
--limit 100 · 61 subtasks
↓0.0000.000
−3.3 ptGSM8K
flexible-extract
GSM8K
flexible-extract
↓0.000.00
limit-noiseKL divergence
vs base
KL divergence
vs base
↑0.000.00
SFT-expected↳ Unbound unbound-e4b fine-tune of gemma-4-e4b-it. AdvBench 520 evaluated with an LLM-judge harness.
Unbound · research previewevalengine.ai ↗