Sparrow Decider
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity5.6/10
Apache-2.0 models on HF, pip package (numpy only), C99 runtime, ESP32-S3 and Strands-tool guides, research report. Anonymous community project, explicitly not AWS/Strands: no support or SLA.
- Capability7/10
No board row. Noul / Choice / Score with per-kind temperature, no ECE. Author: 64 ms p50 (small) to 0.06 ms (nano) on one laptop core; JevBench proxy 122/231; noul near chance at every tier.
- Adoption0/100
Created 23 Oct: 0 HF downloads and 0 likes on the six models, new GitHub repo. No integrations or users found.
Vendor claims
- 9.3 µs per decision on an ESP32-S3 (sparrow-pico); nano 3.5–3.8 ms on ESP32-S3; small 64 ms p50 on one Apple M4 Pro core[Vendor claim — not independently verified]
- sparrow-small-general: JevBench public 122/231 (local v1.5 proxy), MASSIVE domain/intent 0.673/0.599 vs Strands Decider 2B 0.86/0.871, RobotBench 24/55[Vendor claim — not independently verified]
publicado (OK do Fabricio 24/10 BRT). Scores are calculated from scratch.
What it is
Sparrow Decider (GitHub, Hugging Face, Apache-2.0, 23 Oct 2026) is a family of decision models small enough for an agent, a robot or a microcontroller. They answer yes/no (noul), pick-one (choice) and rate-on-a-scale (score) questions in the same shapes as Strands Decider, from which the encoder tiers are distilled (a Strands Decider 4B teacher, 50% teacher / 50% gold). “Unofficial community project. Not an AWS or Strands product.”
| Tier | Model | Size | Input | Runs on |
|---|---|---|---|---|
| pico | compiled per device (e.g. sparrow-pico-rover) |
~2 KB / question | numbers, flags, enums | any MCU |
| nano | sparrow-nano-general, sparrow-nano-robot |
2.2 MB / 1.11 MB | text | ESP32-S3, Pi, laptop |
| micro | sparrow-micro-general |
7.6 MB | text | Pi, laptop, 16 MB ESP32-S3 |
| tiny | sparrow-tiny-general |
17.9 MB | text | Pi, laptop |
| small | sparrow-small-general (Ettin 68M encoder + pointer head) |
47 MB | text | Pi, laptop, Jetson |
pip install sparrow-decider (numpy only); sparrow-decider serve exposes POST /v1/systemone; a C99 runtime with no malloc runs the small tiers on chips.
Limits
- No board result. Every number is the project’s own local run.
- Weak yes/no: the card says
noulon JevBench is near chance at every tier; a value near 0.5 means “don’t know”. - Far behind the 2B teacher on open language (MASSIVE domain 0.673 vs 0.86). English only.
- The authors say never to use it as the only safety stop on a physical system.
Author claims (UnverifiedClaim)
- 9.3 µs per decision on a $10 ESP32-S3 (pico); nano 3.5–3.8 ms on ESP32-S3; small 64 ms p50 on one laptop core (repo).
- small: JevBench public 122/231 (local v1.5 proxy), RobotBench 24/55 (nano-robot 36/55) (card).
Scores (from scratch, 24 Oct 2026)
- Capability 7.0 = types 10 × 0.30 + calibration 4 × 0.25 + latency 8 × 0.25 + accuracy 5 × 0.20 = 3.00 + 1.00 + 2.00 + 1.00 = 7.00. Types 10: noul/choice/score. Calibration 4: per-kind temperature, no ECE. Latency 8: well under 100 ms on CPU (author). Accuracy 5: no board, author numbers in public files (rule of 20/10).
- Maturity 5.6 = availability 8 × 0.30 + docs 7 × 0.25 + integrations 4 × 0.25 + support 2 × 0.20 = 2.40 + 1.75 + 1.00 + 0.40 = 5.55 → 5.6.
- Adoption 0 = engagement 0 × 0.35 + ecosystem 0 × 0.35 + production 0 × 0.30 = 0.