Kestrel Robotics · Lab 04

Smaller is useful.
Qualified is deployable.

Check whether an edge runtime fits your task before you commit. Compare quality, latency, memory and power in one review.

Interactive concept · Fictional task and benchmark samples. No models run, convert or deploy in this demo.

01 / Evaluation setup

02 / COMPARE & QUALIFYNOT EVALUATED

Your constraints.
Your decision.

Run a sample comparison to expose the trade-offs.
Quality is a gate, not a footnote.

F16 reference

52 ms

I2_S fast

Same task / same CPU	F16	Candidate	Gate

Show sample method and provenance

Checkpoint: native BitNet embedding 270M
Task set: kestrel-retrieval-02 · 2,000 queries
Batch: 1 · sequence: 256 tokens · threads: 8
Warm-up: 20 iterations · timed: 200 iterations
Quality: retrieval success @ 5 · reference: 97.2%
Memory: peak process RSS · power: whole-system sensor
Fixture: sample-kestrel-20261002 · no real measurements
Same native checkpoint in F16 and I2_S encodings. Results do not compare different pretrained architectures.

I reviewed the task, constraints and sample evidence. This is a test decision, not production approval.

03 / Qualified evaluation concept

One model. One target. One reviewable pack.

Proposed $2,500 per evaluation: scoped compatibility review, task comparison and reproducibility pack. No guaranteed conversion or speedup.

I understand this purchase is simulated; no payment or service is created.

Sample pricing · one project

Smaller is useful.Qualified is deployable.

01 / Evaluation setup

Kestrel · qualification record

One model. One target. One reviewable pack.

Smaller is useful.
Qualified is deployable.