---
pretty_name: Ralph sub2 physical-iPhone device lab
license: other
language:
  - en
tags:
  - on-device-llm
  - iphone
  - gguf
  - llama-cpp
  - metal
  - benchmark
configs:
  - config_name: default
    data_files:
      - split: test
        path: matrix.csv
---

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

**One configured 2.94 GB Round 7 sub2 crown. Four physical iPhones. Four open
receipts.** Each phone loaded the model and completed the same normal PocketPal
chat prompt through the local llama.cpp Metal runtime.

![Ralph sub2 results across four physical iPhones](social/matrix-hero-1600x900.png)

[Watch the 30-second desktop cut](social/ralph-sub2-four-physical-iphones-horizontal.mp4)
· [Watch the 30-second vertical cut](social/ralph-sub2-four-physical-iphones-vertical.mp4)
· [Watch the 15-second vertical teaser](social/ralph-sub2-four-physical-iphones-teaser-15s.mp4)
· [Download the model](https://huggingface.co/RalphLabsAI/ralph-crowns/blob/main/ralph-qwen3-8b-sub2.gguf)
· [Inspect the machine-readable matrix](matrix.json)
· [Verify every file](SHA256SUMS)
· [Read licensing and attribution](ATTRIBUTION.md)

## Normal PocketPal chat

| Physical device | OS | Chat speed | TTFT | Result |
|---|---|---:|---:|---|
| iPhone 15 | iOS 26.6 | 10.58 tok/s | 838 ms | Completed |
| iPhone 16 | iOS 18.6 | 12.36 tok/s | 439 ms | Completed |
| iPhone 17 | iOS 26.6 | 16.77 tok/s | 427 ms | Completed |
| iPhone 17 Pro Max | iOS 26.6 | 18.54 tok/s | 400 ms | Completed |

These values come from PocketPal's normal-chat timing footer. They are separate
from the controlled `pp` and `tg` benchmark values preserved in `matrix.json`
and must not be substituted for them.

The fixed prompt was:

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

All four responses, timing footers, benchmark screenshots, and untouched
device-screen recordings are included. PocketPal displayed **Memory tight** on
all four devices; the warning is intentionally visible in the evidence.

## Artifact identity

| Field | Value |
|---|---|
| Repository | `RalphLabsAI/ralph-crowns` |
| Filename | `ralph-qwen3-8b-sub2.gguf` |
| Size | 2,937,263,168 bytes (2.94 GB decimal) |
| Expected SHA-256 | `9c45bf0ac486e793bd761a99ce4b9d965ec12e70f2d420e02af16797c71f6b50` |
| Crown | Ralph Round 7 sub2 |
| Profile | Q2_K-led mixed quantization |
| Winning builder | [boweizh1204](https://huggingface.co/boweizh1204) |
| Source artifact | [`boweizh1204/ralph-qwen3-8b-sub2-qat5`](https://huggingface.co/boweizh1204/ralph-qwen3-8b-sub2-qat5) |

The expected digest was recomputed against Ralph's local mirror while preparing
the test build. It was **not computed on-device**. The model weights are not
redistributed in this dataset; download them from the linked model repository.

## What was tested

- BrowserStack App Automate physical iPhones, one cloud-device session per row
- PocketPal 1.17.3 (build 147), test-only device-lab build
- PocketPal source commit `6683a28876d031484cc72dde251b6139a804ad64`
- Context 2,048; GPU backend requested; Metal execution observed as `MTL0`
- Controlled benchmark: pp=128, tg=32, parallel=1, repeats=3, mmap enabled
- Network connected throughout each run

`PHYSICAL_DEVICE_PROOF.md` records how the selected App Automate catalog entries
were checked as physical devices. Each directory under `devices/` contains a
human-readable receipt, three screenshots, and the untouched screen recording.
The `social/` directory contains clearly identified edits derived from those
receipts.

## Claim boundary

These runs show that the configured Ralph sub2 model loaded and generated
through PocketPal's local llama.cpp Metal runtime on the four recorded physical
cloud devices, at the displayed one-session speeds.

They do **not** establish:

- App Store memory-fit certification or required device RAM;
- airplane-mode or offline operation;
- universal iPhone compatibility or typical repeated-run performance;
- an on-device SHA-256 match; or
- broad model quality from one fixed prompt.

BrowserStack re-signing strips PocketPal's increased-memory-limit and
extended-virtual-addressing entitlements. The devices remained connected, the
runs were not thermally normalized, and observed peak memory is a harness
observation rather than a hardware requirement.

“The same model” means the same admitted Qwen3-8B architecture and parameter
structure—not identical weights, guaranteed identical behavior, or universal
phone compatibility.

## Verify the receipt set

```bash
hf download RalphLabsAI/ralph-device-lab \
  --repo-type dataset \
  --local-dir ralph-device-lab
cd ralph-device-lab
sha256sum -c SHA256SUMS
```

The publisher regenerates `SHA256SUMS` from the exact public payload and commits
the card, records, media, and integrity manifest under one protected Hub parent.

## Build on the evidence

Try the same artifact and report both successes and failures. Include your
device, OS, PocketPal version, exact filename, context, load result, timing
footer if shown, and a screenshot or recording. Do not include account details,
serial numbers, notifications, or private device information.

Ralph is an open model-compression tournament on Bittensor SN40. The model card,
signed round record, licenses, notices, and builder credit remain with the
[Ralph crowns repository](https://huggingface.co/RalphLabsAI/ralph-crowns).