Intention
Classifiers that read a conversation window, or a scene sketch plus an observed track, and return labels a host uses to allow, steer or abort its next step.
Falcon · Ducky Software
Models
Falcon is a family of four series of small, specialised models built by Ducky Software: LIM (intention), Waymark (spatial), Express (speech and sound) and Response (short replies).
5 LIM models3 Waymark models3 Express models1 Response models6 families12 releasedupdated
Series · Five models · One family
LIM
Intention
LIM is a series of five small intention models built by Ducky Software.
LIM models in series order
ELIM
Reads the last few lines of a chat and says whether a reply is fine to give, needs a nudge, or should stop.
Conversation-intention classifier: trajectory, allow / steer / abort, harm and steer hint
ELIM reads the last three turns of a conversation and returns four labels in under a millisecond: trajectory, allow / steer / abort, harm class and steer hint.
v1.0.0
LIM
The standard chat safety check: reads the last few lines and says whether the reply is fine, needs a nudge, or should stop.
The standard conversation-intention classifier on a single hashed feature bag
LIM is an 8.9M-parameter intention classifier with the same four heads and decision rules as ELIM on a single-bag backbone, selected with model: lim.
v1.0.0
LIM Nano
The smallest version of the chat safety check, for when every bit of memory counts.
Smallest conversation-intention classifier: the original 256-wide single-bag model
LIM Nano is a 4.3M-parameter intention classifier with the same four heads, vocabularies and decision rules as LIM, at 256 width and half the weight file.
v1.0.0
LIM3D
Watches how someone moves through a photo and tells you where they’re going and if they should stop.
Intention from a scene sketch and an observed track
LIM3D reads a scene sketch and the first seconds of an observed track and returns the mover’s intent, phase, next hazard, proceed / caution / stop and ETA.
v1.0.0
LIM3D-XL
The more careful version of the movement reader: it looks at every object and every step before it answers.
Attention over every object and every step of an observed motion
LIM3D-XL is a 36.2M-parameter transformer that reads a scene sketch and an observed track and returns intent, phase, hazard, proceed / caution / stop and ETA.
v1.0.0
Series · Three models · One family
Waymark
Spatial
Waymark is a series of three small spatial models built by Ducky Software.
Spatial
Models that read a scene and say where things are: a Falcon scene sketch for a layout answer or a timed walk in the viewer frame, or an airspace scene and a flight track for readings and routes.
Waymark models in series order
Waymark
Give it a written sketch of a photo and ask where something is; it answers in one short sentence.
Sketch-to-layout language model for viewer-relative scene answers
Waymark reads a text sketch of a scene and answers one layout question — left or right, closer or farther, in front or behind — in one short sentence.
v1.1.0
Waymark Extra
Give it a sketch of a scene and a goal like “walk to the door”, and it describes the walk step by step.
Movement traces from a scene sketch and a goal, with a plain-words summary
Waymark Extra is a 355M-parameter decoder-only model that turns a scene sketch and a movement goal into a timed, viewer-relative trace of up to 12 steps.
v1.0.0
Waymark Flight
Describe the airspace around a flight and where the aircraft has been, and it says what it is inside, what comes next and how to go around.
Airspace readings and route proposals from a scene and a flight track
Waymark Flight is a 21.3M-parameter transformer that reads airspace and a flight track, says where the aircraft stands and proposes routes a checker validates.
v1.0.0
Series · Three models · Three families
Express
Speech and Sound
Express is a series of three small speech and sound models built by Ducky Software.
Voice
A text-to-speech model that reads a short line of text and returns a spoken WAV in one of seven fixed male and female voices, with no cloning.
Sound
Models that read a short description of a sound and drive a deterministic synthesis engine to render a cue of at most a few seconds: tones, chimes, alerts, textures.
Transcription
A model that turns a short English recording into lower-case text with no punctuation, trained from scratch on read audiobook speech.
Express models in series order
Express Cue
Describe a sound in words and get a short audio clip back.
Renders a short sound cue from a text description through a deterministic synthesis engine
Express Cue reads a short description of a sound and returns a rendered cue as 24 kHz mono WAV: a kind and synthesis parameters that a fixed engine renders.
v1.0.0
Express Voice
Turns a short line of English text into speech, in a choice of clear male and female voices.
Text to speech in seven English voices, male and female: two SpeechT5 fine-tunes
Express Voice speaks up to 600 characters of English in one of seven male and female voices and returns 16 kHz mono WAV: SpeechT5 fine-tunes on two corpora.
v1.1.0
Express2Text
Send a short recording of someone speaking English and get the words back as text.
Turns a short English recording into plain text, trained from scratch on read speech
Express2Text reads a short English recording — WAV, FLAC or OGG, up to a minute — and returns the words as lower-case text without punctuation.
v1.0.0
Series · One model · One family
Response
Response
Response is a series of one small reply model built by Ducky Software.
Response
Small models that write or select a short reply for a host that needs an answer without calling a large responder.
Response models in series order
Cognitio
Writes a short, friendly reply to small talk without calling a big language model.
Short conversational replies from a small fine-tuned instruct model
Cognitio is an 18.5M-parameter LoRA adapter on a 1.5B open-weight instruct model that returns a one- or two-sentence reply to a short window of turns.
v1.0.0
- Released Weights are final and a hosted endpoint serves this version.
- Trained Weights are final; a hosted endpoint for this version is not yet available.
- Training A training job is in progress; weights and metrics are not final.
- Planned Specified but not yet trained.
- Deprecated Superseded; the endpoint may be withdrawn after the notice period on the changelog.