LFM2.5 350M on iPhone 12 · 4 GB

Experiment Complete Published

ExecuTorch 1.6.0 on iPhone 12 · 4 GB, CPU

LFM2.5 350M as 8da4w (4-bit), archive 2048 input / 128 output: a 2048-token prompt and 128 tokens out. Measured 13 Nov 2026, 02:30 UTC by Glenn Sonna, Xybrid.

Doesn’t count in rankings: Its output check failed: 5/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128.

Experiment id s4-iphone12-lfm2.5-350m-4-bit-et-xnnpack-2048

Summary

Over 6 valid trials of 6 attempted (6 planned), computed from the raw trials below.

MetricMedianRangeMean ± SDQuartiles
Decode speed tok/s, higher is better83.5 81.0–84.0 83.2 ± 1.1 83.4–83.6
Time to first token ms, lower is better4,731 4,723–5,197 4,813 ± 189 4,728–4,760
Peak memory GiB, lower is betterNot measured
Prefill speed tok/s, higher is better432.8 394.1–433.6 426.0 ± 15.7 430.2–433.1

Trials, metric by metric

The line runs from the lowest to the highest valid trial, the dot is the median and each tick is one trial.

Decode speed tok/s, higher is better
Time to first token ms, lower is better
Peak memory GiB, lower is better
Prefill speed tok/s, higher is better

Every trial

Raw token counts and durations, with the speeds computed from them. Invalid trials stay in the record with their reason; they are left out of the summary, never silently dropped.

Every attempted trial
TrialTime to first token msPrefill tok/sDecode tok/sPeak memory GiBThermalNote
15,197394.12,048 tok in 5,197 ms, engine81.0127 tok in 1,568 ms, enginenot recordednominal → nominal
24,769429.42,048 tok in 4,769 ms, engine83.7127 tok in 1,518 ms, enginenot recordednominal → nominal
34,723433.62,048 tok in 4,723 ms, engine83.4127 tok in 1,522 ms, enginenot recordednominal → nominal
44,728433.22,048 tok in 4,728 ms, engine83.4127 tok in 1,522 ms, enginenot recordednominal → nominal
54,734432.62,048 tok in 4,734 ms, engine84.0127 tok in 1,512 ms, enginenot recordednominal → nominal
64,728433.12,048 tok in 4,729 ms, engine83.5127 tok in 1,521 ms, enginenot recordednominal → nominal

Caveats

  • Publisher cards identify the base model but do not always pin the exact export source checkpoint revision.
  • Engines may differ in quantization scheme, effective KV cache precision, kernels, and activations.

Configuration

Device

Device
Apple iPhone13,2
Cohort
iPhones (mobile)
SoC
Apple A14 Bionic
CPU
6-core CPU (2 performance, 4 efficiency)
GPU
4-core Apple GPU
NPU
16-core Neural Engine
RAM
4.00 GiB
OS
iOS 27.2.1 (24B156)
Kernel
Darwin Kernel Version 27.2.0: Tue Oct 13 21:44:36 PDT 2026; root:xnu-13432.42.3~28/RELEASE_ARM64_T8101
Power
usb
Low power mode
No
App state
active

Engine

Engine
ExecuTorch
Maintainer
PyTorch (Meta)
License
BSD-3-Clause, open source
Version
1.6.0
Commit
4776f71b6f0daa5dfc7d758d5836fc1cc0d40154
Binary SHA-256
c2ad72f6c7e394f758cbc77dc71c868a9c3d52371bc1379a9e49e3940bc22586
Built from
package
Build flags
None
Backend
CPU
Threads
Not recorded
Interface
ios-app
Settings in effect
backend=cpu app=XybridBench process=one app launch per generation threads=engine default (its threadpool) recipe=lfm2.5-350m-et-xnnpack-8da4w,lfm2_xnnpack_q8da4w.yaml,lfm2.5-350m-et-xnnpack-8da4w.pte echo=false temperature=0 num_bos=0 num_eos=0 reset_before_generation=true

Model and weights

Base model
LFM2.5 350M (LFM2.5, 350M)
Upstream
LiquidAI/LFM2.5-350M@9921cd4d82c0e3986887aa4b376f9446729ef8a3
Weights
LiquidAI/LFM2.5-350M@9921cd4d82c0e3986887aa4b376f9446729ef8a3
Format
pte
File 1
lfm2.5-350m-et-xnnpack-8da4w.pte · 434.9 MiB · sha256 9742351692899c943bf1016142451cd048ead5031b5bea0d3b4ad54b355cdecd
File 2
tokenizer.json · 4.5 MiB · sha256 df1d8d5ec5d091b460562ffd545e4a5e91d17d4a0db7ebe733be34ed374377bd
Quantization
8da4w (4-bit)
Weight bits
4
Group size
Not recorded
KV cache
fp32
Conversion
Converted by Xybrid with executorch export_llm 1.6.0
Conversion command
cd "$ROOT/engines/executorch/build/lfm2.5-350m-et-xnnpack-8da4w" && PYTHONPATH="$ROOT/engines/executorch/.venv/lib/python3.12/site-packages/executorch" "$ROOT/engines/executorch/.venv/bin/python" -m executorch.extension.llm.export.export_llm --config "$ROOT/engines/executorch/.venv/lib/python3.12/site-packages/executorch/examples/models/lfm2/config/lfm2_xnnpack_q8da4w.yaml" +base.model_class=lfm2_5_350m "+base.checkpoint=$ROOT/engines/executorch/build/lfm2.5-350m-canonical.pth" "+base.params=$ROOT/downloads/executorch/4776f71b6f0daa5dfc7d758d5836fc1cc0d40154/examples/models/lfm2/config/lfm2_5_350m_config.json" +export.max_seq_length=4096 +export.max_context_length=4096 +export.output_name=lfm2.5-350m-et-xnnpack-8da4w.pte

Workload

Workload
Archive 2048 input / 128 output
Prompt
2,048 tokens, from prompts/source.txt
Prompt SHA-256
c10d1b7bef62d95322f8f8d229dc66118d9375107bdc49b06f76287d058b2f6c
Chat template
ba551d58630afa3190b1be3602e28301f3d2e9bbac978dfc49d6d825171648b6
Output
128 tokens, ignoring end of sequence
Context length
4,096
Batch size
1
Sampling
Greedy
Timing
One streamed request per trial
Warmup
2 runs, not counted
Trials planned
6
Cooldown
60 s between trials

Conditions

Ambient
Not recorded
Thermal at start
nominal
Thermal at end
nominal
Thermal gate
iOS thermal state nominal
Throttling detected
No

Provenance

Measured by
Glenn Sonna, Xybrid
Operator affiliation
Xybrid
Run at
13 Nov 2026, 02:30 UTC
Harness
xybrid-bench 0.4.0 (8130262155c269b3ee57ab53b9fb646fd29790c2)
Methodology
v0.6
Publication
published
Published
16 Nov 2026, 10:23 UTC
Reviewed by
Glenn Sonna

Output check

Result
Failed
Method
sanity-v1-bf16-reference
Reference
checks/sanity-v1.json
Reference SHA-256
a702dee079525ae0ec90465b1c9e26052c99cb3e1c53e83dea05be4c194fc7ca
Required passing checks
7
Check arithmetic
Failed · sha256 e629fa6598d732768f7c726b4b621285f9c3b85303900aa912017db7617d8bdb
Check capital
Passed · sha256 5dd272b4f316b776a7b8e3d0894b37e1e42be3d5d3b204b8a5836cc50597a6b1
Check opposite
Failed · sha256 0b0fd31e2478cff0d3c5bf526a0f8f1c780abfb435eec0a8f2c9beed0eaa6bc4
Check weekday
Passed · sha256 7d8af1de1262f150187d5938bf649d6c35726970685f916d1ec6392a801b9762
Check color
Failed · sha256 ec7d56a01607001e6401366417c5e2eb00ffa0df17ca1a9a831e0b32c8f11bf7
Check sort
Failed · sha256 0ca7eabde70df147ffea0ec2627afd6c362a0bfc80b48981b8880485cca2360e
Check animal
Failed · sha256 d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35
Check word
Failed · sha256 f61e68691bdd0f5ebd3252de0396569d624bd2fba240d66fae90edb456a9f48f
Check copy
Passed · sha256 d61e6890d0e3571b0c7693612d41f8d059a1fcdef4ba959c0af4144541d41907
Check needle-start
Passed · sha256 8b9af2e5b8b4cfddb87d171997b95c471f8d68a68e30fb5177abeb3332372edd
Check needle-end
Passed · sha256 e6589c736af32f81bf239c5eda6de41959122ccc1d05463b75e0eb0c70fc831c
Check continuation-128
Passed · sha256 46e2ace9e836d0de7961085acc08170024e02dee11c4040d0a28e45dade9f606
Detail
5/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128.

Reproduce it

Commands
pnpm bench run --plan plans/s4-iphone12-lfm2.5-350m-4-bit.json

Three palindromic blocks per quantization class and workload; load, two warmups, one trial, unload per turn.

Raw data