RalphLabsAI

/

ralph-crowns

Like 1

Follow

RalphLabs AI 3

The same model. Small enough to sit on your phone.

We ran the configured Round 7 sub2 crown in PocketPal on four physical iPhones.
Each loaded and answered the same prompt through the local llama.cpp Metal runtime.
These are Ralph team-authored tests, not independent community reports. Now
we'd like to see what happens on your own hardware—successes and failures welcome.

Watch the 30-second edit
· All uncut receipts, screenshots, settings, and checksums
· Website run guide

Physical BrowserStack iPhones; PocketPal test-only build; network connected.
All four displayed Memory tight. These runs are not App Store memory-fit
certification or an offline test. Full limitations are below and in the receipts.

What we recorded

Normal PocketPal chat timings, one session per device, recorded September 15, 2026:

These are the normal-chat footer readings, not the separate controlled
benchmark measurements in matrix.json.

Try the exact file

Download the tested Round 7 sub2 GGUF — 2.94 GB

• Filename: ralph-qwen3-8b-sub2.gguf
• Exact size: 2,937,263,168 bytes
• SHA-256: 9c45bf0ac486e793bd761a99ce4b9d965ec12e70f2d420e02af16797c71f6b50
• Builder: boweizh1204
• Source artifact
  · Apache-2.0 license and preserved notices

In PocketPal, search Hugging Face for RalphLabsAI, open ralph-crowns,
and select that filename. If the current listing changes, use the pinned download
above and import it as a local GGUF. Record your installed app version—our test
build is not the public App Store distribution, and your result may differ.

Our context was 2,048 tokens. Use the same prompt if you'd like a comparable starting point:

In two sentences, explain how lower-bit weights reduce memory use without changing a model's architecture.

Reply with your receipt

You can use the Pocket Lab receipt builder
or copy this template. Please report your actual setup; other phones, desktops,
and runners are welcome too. A failed load is a useful result.

Device / OS:
App + exact version (App Store, source build, etc.):
Exact GGUF filename + revision:
SHA-256 checked: not checked / another machine / on-device
Context / GPU offload / threads, if shown:
Load state: reboot-cold / app reload / cached
Loaded: yes / no
Memory warning, if any:
Prompt, if different:
Normal-chat tokens/s / first-token time, if shown:
Result or exact failure:
Network state / airplane-mode rerun, if tested:
Screenshot or recording link:
May Ralph quote/repost this receipt with your handle? yes / no

Remove notifications, account details, serial numbers, and other private
information before posting. A reply does not grant us permission to reuse your
media; say whether you want it featured and credited.

What these tests do—and don't—establish

• PocketPal 1.17.3 (build 147), test-only device-lab build; source
  6683a28876d031484cc72dde251b6139a804ad64. Local Metal execution was observed as MTL0.
• BrowserStack re-signing strips the increased-memory-limit and extended-virtual-addressing
  entitlements. The memory warning stayed visible in the evidence.
• Network stayed connected. We have not established airplane-mode or offline operation here.
• The expected model hash was verified against Ralph's local mirror, not computed on-device.
• One session per device, without thermal normalization: not typical repeated-run performance,
  universal compatibility, required RAM, or a broad model-quality result.
• This receipt set covers only this exact sub2 artifact—not the other tiers or future replacements.

“Same model” means the same admitted Qwen3-8B architecture and parameter
structure, not identical weights or guaranteed identical behavior. Ralph's
protocol retention is separate from task capability and phone speed.

The crown was built by boweizh1204 from Qwen3-8B; Ralph provides the verified
mirror, tournament record, and these device tests. PocketPal is an independent
project and is not endorsing Ralph. Media terms and attribution
· How Round 7 selected the crown