Families

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). The twelve models fall into five families by the kind of input they read and the kind of answer they give; three series hold one family each and Express holds two. Family is a fixed property of a model, and every family belongs to one series. Role — Default, Alternate or Reserve — describes a model's position among its siblings, and status describes where its current version is in its lifecycle. Series, family, role and status are independent of one another.

LIM series #

LIM is a series of five small intention models built by Ducky Software, all in the Intention family: three read a conversation window and two read a scene sketch plus an observed track.

Intention #

In plain terms: these models look at what has happened so far and say what should happen next.

Intention models answer one question: given what has been observed so far, where is this heading, and what should the host application do about it? Three members read text; two read a scene and a track. The text members share the /v1/intention route and are selected per request with the model field — "elim" (the default), "lim" or "lim-nano" — and the response names the one that scored in result.model.

ELIM (Enhanced Large Intention Model) is the Default text member. It reads a conversation window — up to three prior turns plus the current one — through two hashed n-gram bags and returns four labels: a trajectory (one of 32), an allow / steer / abort action, a harm class and a steer hint, in under a millisecond on a laptop CPU, with a suggested short reply on steer and abort. The host runs it before its responder and acts on the action.

LIM (Large Intention Model) is the standard model of the series and the one it is named after: the same window, the same four heads, the same decision rules and the same response shape as ELIM, on a shallower backbone with one shared 512-wide bag instead of two. A host selects it per request with model: "lim"; the reference runtime also falls back to it when the ELIM weights are absent. ELIM stays the default on the route.

LIM Nano is the original 256-wide LIM, kept as the smallest option in the family: the same hashed bag, the same two GELU layers and the same four heads as LIM at half the embedding width — 4,338,995 parameters and 17.4 MB of weights against LIM’s 8,940,083 and 35.8 MB. It loads from its own bundle beside the other two and is selected with model: "lim-nano"; it has served on the hosted route since 2026-09-13.

LIM3D extends the idea to space. It reads a scene sketch and the first seconds of an observed track in the viewer frame and predicts, for the mover, an intent, a phase, the next hazard, a proceed / caution / stop reading, advice, an endpoint and an ETA. The track can come from a tracker or from a Waymark Extra trace. It is a text bag plus a motion MLP plus an object set encoder — 16.5M parameters, single-digit milliseconds per read in-process — and the low-latency Alternate on the hosted route, selected per request with model: "lim3d"; the response then carries a warning that it identifies the target less often than LIM3D-XL and degrades when steps are missing.

LIM3D-XL enhances LIM3D. Same input contract, same eight readings, same runtime and decision rules, but a six-layer transformer over one global token, up to 12 step tokens and up to 16 object tokens, trained on 1.4M examples with tracker-style noise (position jitter and dropped steps). It is 36.2M parameters and costs about 100 ms per read in the reference runtime (6.8 ms as a batch-1 forward pass in PyTorch), and it is the Default behind /v1/intention3d: a request without a model field is answered by LIM3D-XL.

Waymark series #

Waymark is a series of three small spatial models built by Ducky Software, all in the Spatial family: one answers a layout question, one walks a person through the scene, and Waymark Flight reads an airspace scene and a flight track.

Spatial #

In plain terms: these models answer questions about a scene, or describe a walk through it, from a written sketch — or, for Waymark Flight, say where an aircraft stands in a described airspace and how to route around what it must avoid.

Spatial models never see pixels. Waymark and Waymark Extra read a scene sketch — structured text a vision model produces from an image: named objects, each with a lateral band and a depth band, plus the relations between them — and reason about layout in the viewer's own terms: left or right, closer or farther, in front or behind, what is blocking what.

Waymark answers a layout question about a sketch in one short sentence. It is a 124M-parameter decoder-only transformer with a 512-token window, trained on 2,275 synthetic sketch-and-question rows drawn from 45 object names, 10 viewer phrases and 32 question templates. Answers use relative vocabulary only — no metres, no bearings. It is the Default of the family and the model behind /v1/spatial.

Waymark Extra takes a goal instead of a question (“walk to the doorway”, “the dog walker walks toward you”) and emits a timed trace of up to twelve steps — one line per step with approximate x, y and z in the viewer frame, a heading and notes such as sidestepping an obstacle, giving way, stopping at a road edge with no crosswalk, or climbing stairs. It is a 355M-parameter model with 24 layers and a 1,024-token window, trained from scratch on 120,000 movement rows plus the layout rows, so it handles two-leg routes with a waypoint, people walking toward the viewer or off into the background, loitering, and busier scenes with a second person, extra roads and stairs. It runs on a GPU service and is exposed through the Falcon API at /v1/move, which returns the trace, its parsed steps and a plain-words summary.

Waymark Flight reads a different kind of scene: an airspace of up to 24 cylinder or ring-shaped volumes — Class B, C and D surface areas and shelves, restricted and prohibited areas, TFRs and MOAs — plus an observed flight track, in a local east/north frame in nautical miles and feet MSL. It says which volumes the aircraft is in, its ring tier, whether it is under a shelf, the next boundary event and its ETA, whether it will enter a hazard volume, its margins and the ceiling overhead, with an advisory; and it proposes routes of up to eight waypoints around restricted, prohibited and TFR volumes. It is a 21.3M-parameter transformer encoder with a waypoint decoder, trained on synthetic airspace only, and runs on a CPU service behind /v1/airspace/read and /v1/airspace/plan. Every reading comes with the exact geometric check beside it and every served route is validated exactly; Waymark Flight is decision support and education only, not certified for navigation, and knows no real airspace — the caller supplies the volumes.

All four routes count against the spatial quota bucket. Every coordinate Waymark or Waymark Extra emits is reconstructed from band words. Estimates, never measurements.

Express series #

Express is a series of three small speech and sound models built by Ducky Software: one Voice member, one Sound member and one Transcription member, all Released. The Express series page describes the series as a whole, including its base models and licences.

Voice #

In plain terms: these models read a short line of text aloud.

Voice models read a short line of English text — at most 600 characters — and return it spoken as 16 kHz mono WAV in a voice chosen from a small fixed catalogue. The one member is a fine-tune of an open text-to-speech base, SpeechT5, with the HiFi-GAN vocoder, and conditions on a stored speaker x-vector per voice rather than on any audio in the request, so it cannot clone a voice.

Express Voice is the Default and only member: seven consistent voices, male and female, from one GPU service. The default voice lj is fine-tuned on LJSpeech, a public-domain single-speaker corpus, for 12,000 steps; the six voices added on 2026-09-14 are a second fine-tune on six narrators of the Hi-Fi Multi-Speaker English TTS corpus (CC BY 4.0). A request to /v1/speak carries text and, optionally, one of the seven voice ids that GET /v1/voices lists.

Transcription #

In plain terms: this model listens to a short recording and writes down the words.

Transcription models read a short English recording — WAV, FLAC or OGG, at most 60 seconds — and return the words as lower-case text without punctuation. The one member is trained from scratch, with no pre-trained weights, on LibriSpeech, a corpus of read audiobook speech, so it is at its best on clear, read English and weaker on noisy or conversational audio.

Express2Text is the Default and only member: a 20M-parameter conformer with a CTC head over 29 characters, co-hosted on the Express Voice GPU service behind /v1/transcribe. It counts against the audio quota bucket like its siblings.

Sound #

In plain terms: these models turn a description of a sound into a short clip.

Sound models read a short description of a sound and return a rendered cue rather than speech.

Express Cue is the one member. It reads a description of up to 400 characters, predicts a kind from a catalogue of eighteen — tones, chimes, alerts, clicks, transitions, noise textures and a few environmental sounds — and a set of continuous synthesis parameters, and a deterministic signal-processing engine renders them to 24 kHz mono WAV of between 0.1 and 6 seconds. The parameters come back with the audio so a host can edit and re-render them. It is exposed at /v1/sound, where it has served since 2026-09-13, and runs as a CPU service. It never speaks and never plays a recording.

Response series #

Response is a series of two small reply models built by Ducky Software, both in the Response family: one writes a short reply, the other selects a fixed one.

Response #

In plain terms: these models write, or pick, a short reply.

Response models produce the reply itself, after an Intention check has said allow or steer. The family has one member.

Cognitio writes. It is an 18.5M-parameter LoRA adapter on a 1.5B-parameter open-weight instruct base, fine-tuned on about 2,100 short human-written exchanges to answer in one or two conversational sentences. It runs behind its own service contract on a GPU, fronted by the Falcon API at /v1/chat; the host is expected to run an Intention check before calling it.

ERM, a 2.2M-parameter classifier that selected one of 33 fixed replies as the family's Reserve member, was withdrawn on 2026-09-20; /v1/reserve answers 410 route_withdrawn and names /v1/chat as the successor. A host that needs a fallback for an unavailable responder holds a fixed line of its own.

Roles #

Role tags describe a model's position among its siblings and are independent of status and of series. Default is the member a host gets without configuration (ELIM for the text members of the Intention family and LIM3D-XL for its 3D members, Waymark in the Spatial family, Express Voice in the Voice family, Cognitio in the Response family). Alternate is a sibling with the same contract, selected by configuration (LIM, LIM3D, Express Voice Nano, Express Voice Plus). Reserve is kept for a deliberately small fallback for constrained conditions; no current model carries it. The other four — LIM Nano, Waymark Extra, Waymark Flight and Express Cue — carry no role tag; they are distinguished by what they read, not by precedence.

How the series connect #

A typical host runs one series after another: a LIM check on the conversation window, then either a Waymark read — when the turn carries an image with a question or a goal — or a Response call. The Spatial members and the 3D Intention members share one frame, “approx m; x right, y ahead, z up; viewer at origin”, so the steps a /v1/move call returns can be cut after their first entries and handed to LIM3D-XL as the observed track. The home page lists the same models by the need they meet; Input formats specifies the shared sketch, trace and steps formats, and Output vocabularies lists every label each family returns.

The Express series sits after the other three rather than inside them: once a LIM check has allowed a reply and a Response model or the host's responder has written it, Express Voice can speak it, and Express Cue can render the cue that announces it. Nothing in Express reads a sketch or a window; the series meet only at the host.