Models

Pick the smallest bird that does the job.

Every model answers yes/no, pick-one and score questions with calibrated probabilities. They differ in input, size and where they run.

The family

Smallest first. Every field comes from a measurement; fields we have not measured are left out.

pico

sparrow-pico-rover-v1.1

~2 KB

per question (501 parameters)

Numbers in, decision out. You write a spec of sensors, flags, enums, rules and questions, and the compiler writes one tiny MLP per question as a C header. It answers only the questions it was compiled for.

Input: sensor numbers, flags, enums
Runs on: any microcontroller
ESP32-S3: 9.3 µs per decision
Laptop: 0.18 µs
Quality: 94.0% vs the spec's rules (its 2B teacher: 94.2%)

Read the pico guide

nano

sparrow-nano-robot-v1.2

1.11 MB

Reads states like battery: 6% as values, not words, and ships with pip. 3.5 ms per decision on an ESP32-S3.

Input: text with sensor readings
Runs on: ESP32-S3, Pi, laptop
ESP32-S3: 3.5 ms per decision
Laptop: 0.08 ms (C)
Quality: RobotBench 36/55; navigation from sensor numbers 8/10 (public 2B: 3/10)

Read the robot guide Download

nano

sparrow-nano-general-v1.1

2.2 MB

Reads text and answers in 3.8 ms on an ESP32-S3. Good for routing and gating inside an agent when you want no network call.

Input: text
Runs on: ESP32-S3, Raspberry Pi, laptop
ESP32-S3: 3.8 ms per decision
Laptop: 0.06 ms (C)
Quality: general text decisions; JevBench 122/231, MASSIVE intent 0.588

Read the ESP32-S3 guide Download

micro

sparrow-micro-general-v1.1

7.6 MB

More room, more reach: a text model for a Raspberry Pi, a laptop or a Jetson.

Input: text
Runs on: Pi, laptop
Laptop: 0.22 ms (C)

Read the Python guide Download

tiny

sparrow-tiny-general-v1.0

17.9 MB

More room, more reach: a text model for a Raspberry Pi, a laptop or a Jetson.

Input: text
Runs on: laptop, Pi
Laptop: 23 ms (numpy)

Read the Python guide Download

small

sparrow-small-general-v1.0

47 MB

More room, more reach: a text model for a Raspberry Pi, a laptop or a Jetson.

Input: text
Runs on: laptop, Pi, Jetson
Laptop: 64 ms (numpy)

Read the Python guide Download

Compare

The Sparrow Decider family. Timings are measured; blank cells were not measured.
Tier	Model ID	Size	Input	Runs on	ESP32-S3	Laptop	Quality
pico	sparrow-pico-rover-v1.1	~2 KB per question (501 parameters)	sensor numbers, flags, enums	any microcontroller	9.3 µs	0.18 µs	94.0% vs the spec's rules (its 2B teacher: 94.2%)
nano	sparrow-nano-general-v1.1	2.2 MB	text	ESP32-S3, Raspberry Pi, laptop	3.8 ms	0.06 ms (C)	general text decisions; JevBench 122/231, MASSIVE intent 0.588
nano	sparrow-nano-robot-v1.2	1.11 MB	text with sensor readings	ESP32-S3, Pi, laptop	3.5 ms	0.08 ms (C)	RobotBench 36/55; navigation from sensor numbers 8/10 (public 2B: 3/10)
micro	sparrow-micro-general-v1.1	7.6 MB	text	Pi, laptop	not measured	0.22 ms (C)	not measured
tiny	sparrow-tiny-general-v1.0	17.9 MB	text	laptop, Pi	not measured	23 ms (numpy)	not measured
small	sparrow-small-general-v1.0	47 MB	text	laptop, Pi, Jetson	not measured	64 ms (numpy)	not measured

Laptop times are per question on one CPU core, with the runtime in brackets. See Benchmarks for methodology.

Which one?

Your input is sensor numbers and the questions never change → pico
Your input is text and the device is an ESP32-S3 → nano-general
Your input mixes text and sensor readings → nano-robot
You have a Pi, a laptop or a Jetson and want more headroom → micro, tiny or small

Get the models

pip

nano, bundled

numpy only. The package bundles sparrow-nano-general-v1.1 and sparrow-nano-robot-v1.2; load them with Decider.load().

pip install sparrow-decider

Python guide

Hugging Face

micro, tiny, small

Put the .sparrow files in ~/.cache/sparrow-decider/ (or any folder in $SPARROW_MODELS) to load them by name.

Download models

Compile

pico

pico models are compiled for your device from a spec.

sparrow-pico guide

C runtime and ESP-IDF component

Model card

What it does: Answers noul (yes/no), choice (pick one) and score (rate on a scale) questions about a state, with calibrated probabilities, in the same shapes as Strands Decider.
Teachers: A full Strands Decider (4B) teaches the text tiers through distillation on ~430k decisions. pico models are labelled by the public Strands Decider 2B on sampled device states.
Base weights: micro and nano from potion-base-8M (MinishLab, MIT); tiny and small from Ettin encoders (JHU CLSP, MIT).
Runtimes: Python (numpy only); C99, no malloc, model read from flash (ESP-IDF component); sparrow-pico compiler: device spec → C header.
Calibration: Each model ships per-kind temperatures fitted on held-out data. A yes/no answer near 0.5 means "don't know".
Licence: Apache-2.0 (code); model weights derived from MIT-licensed potion-base-8M and Ettin encoders.

Limitations

Open-ended language it never trained on: well behind the 2B (MASSIVE domain routing 0.37-0.67 across released Sparrow models vs 0.86).
Yes/no judgement on JevBench is near chance at every tier.
pico answers only the questions it was compiled for.
These models suggest decisions. Don't use them as the only safety mechanism on a physical system.

Full model card