Can OneJev 4B run on iPhone 15?
Estimated weights fit · custom workflow required
Weight fit is not phone app support
The tracked Q3_K_M weights are 2.5GB; the generic working-memory estimate is 3.5GB against 3.9GB usable. Image inputs need additional projector memory.
OneJev scores typed decisions in a single forward pass, not chat replies. The decode estimates below are not measured decision latency. The publisher's qev server and System One client are required; AICanRun has not verified a phone app route.
Not measured
decision latency
3.5GB
needed at Q3_K_M
Q3_K_M
quant checked
needs 3.5 GBusable 3.9 GB
Apple A16 Bionic6 GB RAMMetal
Make it bearable
1
Drop to Q2_K (2.1 GB) — it runs comfortably here at ~11 tokens/s. Recommended over Q3_K_M.
2
Drop context to 2K — saves ~0.2 GB of KV cache and a little speed.
3
Close every other app before loading — the 0.4 GB headroom is real; iOS may evict the model otherwise.
4
Expect throttling after ~10 min of sustained generation on a phone chassis.
Where the memory goes
| Component | Detail | GB |
|---|---|---|
| Model weights | Q3_K_M GGUF (2.5 GB) + mmap overhead | 2.6 |
| KV cache | 4K context window | 0.3 |
| Runtime | llama.cpp + app overhead | 0.6 |
| Total needed | at Q3_K_M, 4K context | 3.5 |
| Working budget | 6 GB RAM − iOS reserve (jetsam limit) | 3.9 |
| Headroom | remaining inside the working budget | 0.4 |
Pick your quant
| Quant | Download | Verdict | Speed |
|---|---|---|---|
| Q2_K BEST HERE | 2.1 GB | ✓ Runs great | ~11 tokens/s |
| Q3_K_M | 2.5 GB | ! Runs, barely | ~9.2 tokens/s |
| IQ4_XS | 2.9 GB | ✕ Won't fit | won't fit |
| Q4_K_S | 2.9 GB | ✕ Won't fit | won't fit |
| Q4_K_M | 3.1 GB | ✕ Won't fit | won't fit |
| Q5_K_M | 3.5 GB | ✕ Won't fit | won't fit |
| Q6_K | 4 GB | ✕ Won't fit | won't fit |
| Q8_0 | 5.2 GB | ✕ Won't fit | won't fit |
Use the right workflow for this model
The model fits in memory, but it is not an ordinary chat model.
This model is designed to return calibrated probabilities for typed choices from text or images. PocketPal does not provide the required the publisher's qev server and System One client; image inputs also need a vision projector, so the tracked file is not a beginner phone-chat path.
Read the publisher workflow ↗If you only want a normal private chat, use this simpler model instead:
Check Llama 3.2 3BRelated checks
More on iPhone 15
OneJev 4B on other phones