OpenStride

A running coach that reads your body.

OpenStride reads your runs and heart rate from Apple Health, works out your zones and training load, and plans each week. The coach is Gemma 3 4B, an open-weight model running on the iPhone, so your health data stays on the phone.

Gemma 3 4B on-device Apple Health About 20 s per plan on iPhone 13 Pro Works offline

Code on GitHub Watch the demo Sample store listing

What it looks like

A sample runner who doubled their mileage last week. The coach sees a load ratio of 1.98, cuts volume and keeps one hard day.

Weekly plan with a coach note about the load ratio — The week, with a note that matches the plan

Tempo session detail with pace and heart rate targets — Paces and HR zones from your own data

Signals and heart rate zones — Signals with fixed thresholds

The exact prompt the coach saw — Exactly what the coach saw

How it works

Small models are bad at arithmetic and good at reading a situation. So the math lives in code and the judgment lives in the model.

Measure

HR zones (Karvonen), VDOT pace zones, training load, HR drift, last week's completion.

Draft

A plan engine builds the week: phase, volume, long run, quality sessions.

Decide

Gemma reads OK/FLAG signals and picks adjustments from a small action set.

Guard

Rules override anything unsafe. No flag means the draft stands.

Explain

Gemma writes the note from facts about the final plan, so it always matches.

How it is tested

Six synthetic runners, each with an expected coach behavior, run 3 times per model. On-device rows use the same 4-bit MLX weights the iPhone downloads; times on an M2 Pro.

Coach	Final plan right	Model right on its own	Note matches plan	Time
Gemma 3 4B, on-device (MLX)	100%	100%	100%	8.0 s
Gemma 3 4B, Ollama	100%	100%	100%	4.8 s
Gemma 3 1B, on-device (MLX)	100%	33%	100%	4.3 s
Rules only	100%	100%	100%	0 s

Plus 21 unit tests, a simulator UI test that writes a sample runner into Apple Health and reads it back through the real HealthKit path, and a device UI test on a real iPhone.

Why open weights

Private by design

Heart rate and VO2 max are health data. With the model on the phone, the app has no server to send them to.

No per-user cost

Every plan runs on the phone's GPU. No API bill means it can be a one-time purchase.

Works on the trail

After a one-time 2.6 GB download, planning works with no signal.

Inspectable and swappable

The same weights run on a Mac, in the simulator and on the phone, so they can be tested, and a newer Gemma can drop in.