Docs How it works

Results

BOAR's measured results on a real Android phone: answer times, tokens per second, memory and accuracy for five offline models, with the failures and the raw files.

Every number here comes from one phone: a Xiaomi 2311DRK48G (MediaTek Dimensity 8300, 11.6 GB RAM, Android 16), running BOAR's own evaluation. The raw files are in docs/evidence, and you can run the same thing on your phone.

With the Wikipedia pack, the default model

Six knowledge questions with Qwen2.5-1.5B and the Wikipedia Vital Articles pack (2026-09-24):

Measure Result
Answered correctly 6 of 6, each from retrieved articles
Time to the first word 6.4 s on average
Writing speed 17.5 tokens a second on average
Whole answer 10.6 s on average
Peak memory 1.76 GB
Network none

That included questions the built-in library can't answer, like Napoleon's dates.

Five models, 17 questions each

The baseline run, 2026-09-24, same phone, a 512-token answer budget and a 120-second limit per step. Medians:

Model Tokens a second First word Whole answer What happened
Qwen2.5-1.5B (default) 11.4 13.6 s 21.0 s 17 of 17, no timeouts
Adaptive routing 11.4 14.1 s 25 s picks the model per question
LFM2.5-8B-A1B (mixture of experts) 14.8 27.2 s 57.3 s the only model to solve the RAM question; 4 of 17 ran out of answer budget
Phi-3.5-mini 4.0 44.0 s 73.1 s the most complete comparisons; 4 timeouts
Qwen2.5-7B 2.7 70.5 s 107.1 s 5 timeouts
Instella-MoE-16B-A3B – – – didn't load: architecture not supported yet

LFM2.5 ran separately with the same 17 questions. The highest peak memory of any run was 5.21 GB, under the 12 GB limit.

What went wrong

We publish the failures because they're where the work is:

  • Arithmetic with several steps. Every model except LFM2.5 missed the RAM-budget question.
  • Invented citations, especially from Phi-3.5-mini.
  • Timeouts. Nine answers were cut off by the 120-second limit in the baseline.
  • Reasoning that runs out of room. LFM2.5 thinks before it answers, and with a 512-token budget four answers ended before the final line.
  • Heat. Speed fell by about a third over a 66-minute run.

Not measured yet

  • BOAR against internet search plus a frontier model. The bar people care about ("more than half as good") hasn't been tested. We won't claim it until it is.
  • GrapheneOS on real hardware.

Where these numbers come from

The same code runs every evaluation, on the phone or from a computer, and writes one JSON line per answer. The JSONL files are the source of truth; reports and tables are made from them. If our numbers look wrong on your phone, send us yours.

Edit this page on GitHub · Something wrong? Open an issue