Repository navigation
Ralph Device Lab: a 2.94 GB Qwen3-8B GGUF on four physical iPhones (PocketPal receipts) #918
bitzic
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hi PocketPal folks — sharing this on behalf of RalphLabsAI, and thank you for
building PocketPal. We have published a small set of reproducible device-test
receipts that might be useful to people experimenting with larger, lower-bit
GGUFs on phones.
Our Round 7 sub2 crown is a 2.94 GB Qwen3-8B-derived GGUF, built by
boweizh1204. Ralph runs the compression
tournament and publishes the verified mirror; we did not build the winning
quantization ourselves. PocketPal is an independent project, not a Ralph partner
or endorser.
What we tried
The exact same file loaded and answered the same normal-chat prompt on four
physical BrowserStack iPhones. These are our team's tests, not independent
user reports.
Important limitations: PocketPal 1.17.3 (147), test-only build, source
6683a28876d031484cc72dde251b6139a804ad64; context 2,048; local llama.cpp/Metal(
MTL0); network connected. All four displayed “Memory tight.” BrowserStackre-signing stripped the increased-memory-limit and extended-virtual-addressing
entitlements. This is not an App Store memory-fit certification or an offline
test. One session per device, no thermal normalization; these are chat-footer
readings, not the separate controlled benchmark results or repeated-run averages.
We verified the expected hash against our local mirror, not on-device.
Evidence and exact artifact
ralph-qwen3-8b-sub2.gguf, 2,937,263,168 bytes9c45bf0ac486e793bd761a99ce4b9d965ec12e70f2d420e02af16797c71f6b50In PocketPal's Hugging Face model search, look for RalphLabsAI → ralph-crowns →
ralph-qwen3-8b-sub2.gguf. The pinned link above identifies the exact tested file
if the current crown changes. Please use your installed official app release and
report its version; you do not need our test build. Results may differ, including
failure to load. The 2.94 GB download size is not the runtime RAM requirement.
Prompt used:
What would help next
If you already have a phone and want to try it, a single honest report would be
useful. Older iPhones and Android devices are welcome; a failed load or slow run
is useful evidence too. This is optional, unpaid community testing — no purchase,
positive review, or social post is requested.
Reply here, or use our HF Device Lab #1 thread and full receipt template:
We will not copy your report or media into the HF evidence collection or our
campaign without your permission. If there are additional settings that would
make these receipts more useful to PocketPal users, suggestions are welcome.
All reactions