SparseFlow V2 / Independent Apple-device research

Less model.
More possibility.

Can physically compact vision models make on-device AI more efficient? We are investigating that question through Core ML export and real-device iPhone evaluation.

Explore the progress Talk to the founder

Research in progress · Updated October 9, 2026
Independent Maple Silicon work. No Apple partnership, endorsement, or contract is implied.

Core ML / Model evaluation

Dense controlReference model
Channel-pruned baselineComparison model
Physical compactSparseFlow candidate

CONCEPTUAL VIEW
REAL-DEVICE MEASUREMENTS

01 / The research path

From model structure
to device behavior.

A smaller representation only matters if it preserves useful predictions and improves execution on the target hardware. Our workflow examines those questions separately.

Model-level evaluation

Compact and recover

Compare recovered dense, channel-pruned, and physically compact models using a held-out four-class QuickDraw task. Source-model accuracy has been evaluated; that is distinct from on-phone task accuracy.

Core ML workflow

Export and compare

Bring the selected variants into Core ML and check numerical agreement with fixed fixtures. Export and parity checks support the workflow; they do not establish energy or latency gains.

Real-device testing

Measure on iPhone

Run short-session comparisons on iPhone 13 Pro, recording latency distributions and thermal state under CPU-only and all-compute-unit configurations.

02 / Current findings

The result is still an open question.

Short-run latency direction varies across the supplied blocks. Some favor the compact candidate; others show no clear direction. We have not established a general iPhone speedup or a sustained performance advantage.

Research status summary. This is an overview, not a published benchmark release.
QuestionCurrent evidence
Does compaction preserve task quality?Offline held-out source-model evaluation completed for a small four-class task. On-phone task accuracy has not been measured.
Is the compact model faster on iPhone?Mixed short-session latency results; no pooled performance verdict.
Does the benefit persist?Sustained validation is incomplete. The supplied sustained attempt was partial and failed.
Does it save energy?Energy use has not been measured.
Where do the operations execute?Requested compute-unit configuration is recorded. Actual operator placement has not been measured.

The research uses a small vision task and one device family. Findings should not be generalized to other applications or Apple hardware. NVIDIA performance figures elsewhere on this site do not apply to this research.

03 / What comes next

Earn the next
performance claim.

The next milestone is a repeatable, decision-ready device result with clear validation boundaries.

  1. Complete sustained device comparisons with explicit cooldown and thermal conditions.
  2. Measure task accuracy on the exported models on phone, separately from fixture parity.
  3. Investigate actual operator placement and measure energy before claiming battery benefits.
  4. Extend evaluation to additional workloads and devices before making broader claims.

Technical conversations

Working on on-device inference?

Connect with Maple Silicon about model compaction, Core ML evaluation, and research collaboration.

Discuss SparseFlow V2

Founder contact: Gourav Kumar
gourav.kumar@maplesilicon.co