Overfit in the field · iPhone · unlisted
Gemma 4 on iPhone
26B parameters. One phone. No cloud.
A 26B-parameter mixture-of-experts model whose full weights do not fit in an iPhone’s memory, answering a multi-step ballistic-pendulum problem on the device itself. Overfit keeps the shared layers resident and pages routed experts from local storage as the router asks for them — so the model runs at its real size, at storage speed rather than memory speed.
- Prefill speed
- 34.4tk/s
- Prefill time
- 20.34s
- Decode speed
- 3.5tk/s
A ballistic-pendulum problem chained across a spring compression and a pendulum swing. Model loaded and resident at 1.9k / 4.1k context.
1,201 tokens of working, ending at 661 m/s. The run footer reports the on-device timings, and the badges confirm the whole thing stayed local.
The run
Exactly what was on the device.
- Model
- gemma-4-26B-A4B-it — Q4_K_XL
- Parameters
- 26B total · ~4B active per token
- Device
- iPhone — no Mac, no relay, no cloud
- Execution
- Overfit paged experts from local storage
- Context
- 1.9k of 4.1k used
- Output
- 1,201 tokens · 362.45s wall clock
One capture on one device, not a benchmark. Prefill and decode speed depend on storage throughput, available memory, quantization, prompt length, expert reuse, and thermal state — a smaller model that fits entirely in memory will still start and generate faster.
Noema Overfit