Falcon

Falcon · Ducky Software

Changelog

Every version of every Falcon model, plus site-level changes, newest first.

Rows that name a model come from that model's version history and link to its documentation page. Rows marked Site record changes to this site, its four series and its documentation framework. The same list is published as changelog.json.

September 2026

DateModelVersionStatusNote
Waymark1.1.0ReleasedRetrained on a balanced street-and-indoor mix, via an indoor finetune of 1.0.0 (answer-only loss, 1,500 steps, 1× A100). Same contract; one model now covers street and indoor scenes.
Site——ERM withdrawn — ERM is withdrawn the same day it was deprecated: the reserve responder is no longer useful beside Cognitio. Its page redirects to Cognitio, POST /v1/reserve answers 410 route_withdrawn with successor /v1/chat, the reserve quota bucket is gone (four buckets remain: intention, spatial, complete and audio), and the erm field leaves the health report. The Response series has one member, Cognitio; twelve models remain.
Site——ERM deprecated; Cognitio is the successor — ERM, the Reserve member of the Response series, is Deprecated. A generated reply from Cognitio has proved more useful than a fixed line, and a host that needs a fallback for an unavailable responder can hold a line of its own. The page and the weights stay. POST /v1/reserve keeps answering for existing callers until at least 2026-12-19 (the notice period), with deprecated: true and successor: /v1/chat in every response and Deprecation and Sunset headers; it may be withdrawn after that date. ERM is no longer offered on the chat site or the home page. New integrations should call /v1/chat.
Site——Falcon Chat: try every model in the browser — A chat site at chat.falconlab.app lets anyone pick a Falcon model and try it without a credential: a conversation with Cognitio or ERM, a text continuation from the Lumen pilot (the first checkpoint of a from-scratch general language model, 114M parameters), an intention reading from ELIM, LIM or LIM Nano, a layout question to Waymark from a sketch or a photo, and speech, transcription and sound cues from the Express series. The site calls the public playground routes of the Falcon API, /v1/try/<route>, which run the ordinary routes on one shared account with per-visitor limits; the Lumen pilot also answers on POST /v1/complete for keyed callers.
Site——Waymark 1.1.0: one model for street and indoor scenes — Waymark is retrained, via a one-day indoor finetune, on a balanced mix of street and indoor sketches with exact answers (108,825 synthetic rows, 45 % street), with the loss on answer tokens only. On the benchmark's street control set it answers 87 % of questions exactly (1.0.0: 16 %), on hospital scenes 40.5 % (3 %), and in indoor environments it never saw 53 % (4 %). The separate Waymark Indoor page of 2026-09-19 is withdrawn; its request value "model": "waymark-indoor" still works as an alias, and requests and responses are otherwise unchanged. Weight blends of the two models were tried first and rejected because no blend kept both skills.
Express2Text1.0.0ReleasedFirst version: trained from scratch on LibriSpeech in one 3.2-hour A100 job (4 passes); 12.1 % / 25.5 % word error on test-clean / test-other with greedy decoding.
Waymark Flight1.0.0ReleasedFirst released version. Base size, trained on one L4 and early-stopped at epoch 22 of 40; serves behind /v1/airspace/read and /v1/airspace/plan, always paired with the exact checker.
Site——Waymark Indoor: layout answers for rooms, corridors and concourses — Waymark Indoor joins the Waymark series as the Alternate of the layout task: Waymark's 124M-parameter model finetuned on synthetic indoor scenes (homes, offices, schools, shops, restaurants, hotels, station concourses and garages). It answers on /v1/spatial when the request carries "model": "waymark-indoor"; Waymark stays the default, and every /v1/spatial response now names the model that answered in a new model field. On 700 synthetic questions it answers 32.5 % of hospital-scene questions exactly (base Waymark 3.0 %), 46.0 % in indoor environments it never saw (4.0 %) and 33.0 % of Waymark's own street questions (16.0 %). Trained on synthetic data only and not clinically validated; not a safety or wayfinding aid on its own.
Site——Waymark Flight: airspace understanding and route planning — Waymark Flight joins the Waymark series as its third Spatial model: a 21.3M-parameter transformer that reads an airspace scene of up to 24 cylinder or ring-shaped volumes plus a flight track and says which volumes the aircraft is in, its tier, the next boundary event and its ETA, hazard entries, margins, the ceiling overhead and an advisory, and proposes routes of up to eight waypoints. New routes /v1/airspace/read and /v1/airspace/plan (spatial bucket) return every reading beside the exact geometric check and validate every route exactly, falling back to the exact planner. Trained on synthetic airspace only; decision support and education, not certified for navigation. The input formats and vocabularies pages document the airspace scene and Waymark Flight's catalogues.
Site——Hosting pages, separate from the models — Dedicated hosting and the Web Hosting add-on have their own pages — an overview, the compute tiers and Web Hosting — with a Hosting entry in the header; the Dedicated hosting docs topic is retired and its address redirects to the overview. A Web Hosting site is published at https://<site>.falconlab.app, and some site names are reserved for Falcon’s own addresses.
Site——Web Hosting add-on for static sites — An organisation can now add a static site next to its Falcon integration: one static site at its own falconlab.app address with TLS and a global CDN, 1 GB of storage and 20 GB of transfer a month, a custom domain on request. The site name is chosen in the new Web hosting panel of the dashboard, the site is uploaded there as a zip (at most 25 MB, index.html at the top level or in one top-level folder) and a Falcon administrator publishes it within one business day. The add-on is independent of the compute tier and combines with the shared tier or any dedicated tier; billed monthly in CAD, cancel at any time for the end of the month.
Site——Dedicated hosting tiers available on request — An organisation can now rent private, always-warm capacity from the portal: Dedicated Warm (249 CAD a month, a private Falcon API and Express Cue instance with no scale-to-zero wait), Dedicated GPU (1,499 CAD, one NVIDIA L4 kept warm 24/7 hosting one GPU model service: Express Voice, Waymark with Waymark Extra, or Cognitio) and Dedicated GPU ×2 (2,799 CAD, two L4s for up to two of those services, one on each). Requests go from the new Hosting panel of the dashboard to an administrator, who provisions Dedicated Warm within one business day and a GPU tier when GPU capacity is available; once active, Express Cue and the hosted GPU models answer from the organisation's own services with the same keys and routes, and the in-process models run privately at the organisation's private API address. Billed monthly in CAD, cancel at any time for the end of the month; GPU capacity subject to availability, one region (us-central1).
Site——Voice and Waymark move to GPU — Express Voice now serves all seven voices from one Cloud Run service with an NVIDIA L4 GPU, kept warm with one minimum instance; the two CPU services are retired and /v1/speak and /v1/voices are unchanged for callers. Measured through the Falcon API, a 14-word sentence renders in about 2.2–2.5 s (was 11–30 s) and a short line in about 0.7 s. Waymark runs on the same L4 service as Waymark Extra, loaded as a second model behind /v1/spatial: 0.6–1.1 s per answer once warm (was 14–18 s on 2 vCPU), with the service's one-to-two-minute cold start after a quiet period. Waymark Extra and Cognitio run a warm-up at start-up, so the first request after a cold start is no slower than later ones; a twelve-step trace takes about 5–10 s (was 11–14 s) and a Cognitio reply about 0.8 s.
Site——Express Voice: six male and female voices added — Express Voice 1.1.0 serves seven voices behind /v1/speak: the unchanged lj plus female-1, female-2, female-3, male-1, male-2 and male-3, a second SpeechT5 fine-tune on six narrators of the Hi-Fi Multi-Speaker English TTS corpus (CC BY 4.0), 15,000 steps at learning rate 2e-5 on one NVIDIA A100 40 GB, the best of an eight-run sweep. Held-out word error rate 1.6 % across the six, every voice within 1.1–2.1 %, UTMOS 3.76; the new voices read 14- to 20-word sentences intact, unlike the withdrawn multi-voice attempt. The six voices run on a second CPU service, express-voice-duo; the API merges both catalogues on /v1/voices, answers 400 unknown_voice for an id outside the seven and 503 speak_warming with Retry-After: 30 while a service loads.
Express Voice1.1.0ReleasedSix male and female voices added: a second fine-tune of the same base on six Hi-Fi TTS narrators (lr 2e-5, 15,000 steps, held-out loss 0.401) on a second CPU service. The voice lj is unchanged.
Waymark Extra1.0.0ReleasedFirst released version. Trained from scratch on 2 × A100 with early stopping at step 2500; best validation loss 0.166 at step 1750. Serves behind POST /v1/move from 2026-09-13.
Site——Express Voice Nano renamed Express Voice; Express Voice Plus withdrawn — The 12,000-step single-voice model serving as voice lj is now simply Express Voice, and reports engine express-voice; the planned 20,000-step recipe is dropped and the multi-voice Express Voice Plus is withdrawn because its 2026-09-13 evaluation showed the decoder losing its place on sentences longer than about ten words (38.7 % held-out word error rate). Express is a series of two models; the old addresses redirect.
Site——Express Voice Plus trained and evaluated; held back — Two 40,000-step fine-tunes of SpeechT5 on all 109 VCTK speakers finished on one NVIDIA A100 40 GB; the learning-rate 1e-5 run (held-out loss 0.385) is the selected export and the model is now Trained. The 2026-09-13 GPU evaluation records why it is not released: the voice sounds natural (UTMOS 3.71, on par with Express Voice Nano) but the decoder loses its place on sentences longer than about ten words, for a held-out word error rate of 38.7 % against Nano's 1.2 %. The hosted /v1/speak route continues to serve Express Voice Nano only; Voice Plus will be released after a retrain that fixes long-sentence stability, or as a phrase-chunked preview.
Site——Four series: LIM, Waymark, Express and Response — The site now describes Falcon as a family of four series of small, specialised models rather than two. LIM (intention) holds ELIM, LIM, LIM Nano, LIM3D and LIM3D-XL; Waymark (spatial) holds Waymark and Waymark Extra; Express (speech and sound) is unchanged; Cognitio and ERM form the Response series. Families, the models index, the home page and the docs sidebar are grouped by series; every model URL is unchanged.
Site——Waymark Extra (355M) released behind /v1/move — The 355M-parameter movement model, documented until today as Waymark Extra Plus, is now Waymark Extra and serves from a GPU service through the Falcon API at POST /v1/move: a sketch and a goal in, a timed trace, parsed steps and a plain-words summary out, metered in the spatial bucket. Its old address redirects. The 124M preview it supersedes was never published and is retired; the Evaluation chapter records both models on the 2026-09-13 suite (endpoint error 1.84 m against 4.95 m, two-leg goals 25 of 25 against 0).
Site——LIM renamed, LIM Nano released, model selector on /v1/intention — The standard 512-wide intention model is now simply LIM (its slug is unchanged). LIM Nano serves from its own 17.4 MB bundle on the same route. POST /v1/intention takes an optional model field — elim (the default), lim or lim-nano — answers 400 unknown_model for anything else, echoes the model that scored in result.model, and GET /health lists limVariants.
Site——LIM3D-XL, LIM3D and Express Cue released — LIM3D-XL is the default weights behind /v1/intention3d; LIM3D is selectable on the same route with model: "lim3d" and its responses carry a warning about weaker target identification and less robustness to missing steps. Express Cue serves from a CPU service behind /v1/sound. The three Evaluation chapters carry the 2026-09-13 evaluation programme: fresh held-out, robustness, calibration, latency and real-sketch consistency for the two LIM3D sizes; held-out, audio-judge, count and duration results for Express Cue. The site now has seven released models.
Site——Express Voice Nano released: the learning-rate 5e-6 run serves the single voice — The first finished run of the 12,000-step sweep (lr 5e-6, held-out eval loss 0.3689) is version 1.0.0 and serves the voice lj from the Express Voice CPU service behind /v1/speak, with its entry in /v1/voices. The 1e-5 and 2e-5 runs are still training and may replace the checkpoint under the same version if they finish with a lower held-out loss; intelligibility and MOS-proxy results follow when the evaluation job reports.
Site——Express Voice Nano: the 12k-step single-voice recipe, in training — Express Voice Nano is the quicker, cheaper recipe of the Express Voice LJSpeech voice: the same SpeechT5 base, corpus, x-vector and vocoder, fine-tuned for 12,000 steps instead of 20,000, as a three-way learning-rate sweep (5e-6, 1e-5, 2e-5) on Vertex AI started on 2026-09-13; the checkpoint with the lowest held-out loss becomes 0.1.0-pre. It shares Express Voice's catalogue entry and route. Express is now a series of four models.
Site——Express series announced: Express Cue, Express Voice, Express Voice Plus (Planned) — The site then documented two series. Express is a series of three small voice and audio generation models in two families, Voice and Sound, all Planned: Express Cue renders a short sound cue from a description, Express Voice speaks in one fixed voice, and Express Voice Plus speaks in one of twelve documented voices. Three audio routes and an audio quota bucket are specified on the API and Access pages.
Site——LIM Nano listed: the standard 256-wide LIM returns as the tenth model — The original 256-wide Large Intention Model, superseded as the installed LIM weights by the 512-wide LIM retrain, is documented again as LIM Nano, the smallest member of the Intention family. It sits after LIM in series order; the site now describes Falcon as a series of ten models.
Express Voice1.0.0ReleasedThe lr 5e-6 run of the sweep: 12,000 steps, held-out eval loss 0.3689, serving the single voice lj.
Express Cue1.0.0ReleasedFirst released weights: size xl on 2,000,000 synthetic descriptions for 16 epochs on Vertex AI (job express-cue-20260913-092405); evaluated and serving behind /v1/sound since 2026-09-13.
Site——Falcon becomes an independent series; documentation framework and rebuilt site — The site now describes Falcon as a series of nine small, specialised models built by Ducky Software, in three families. Every model has an overview page and a nine-chapter documentation page generated from one source; series topics, a status vocabulary, this changelog and a sitemap are new. Retired self-serve pages redirect to Access.
LIM3D-XL1.0.0ReleasedFirst documented version. Trained from scratch on Vertex AI (job lim3d-xl-20260912-182554), evaluated 2026-09-13, and serving as the default weights behind /v1/intention3d since 2026-09-13.
LIM3D1.0.0ReleasedFirst documented version. Size preset large, trained from scratch on 515,943 synthetic examples (Vertex AI job lim3d-20260912-171944); evaluated and selectable on /v1/intention3d since 2026-09-13.
Site——Self-serve accounts and API keys paused — Sign-up, login and key-minting pages were replaced with archived notices, and sign-up links were removed from the navigation and the home page.
Site——Model pages reduced to short descriptions — Model cards dropped label vocabularies, API routes and pipeline detail in favour of short blurbs; HTML cache headers were hardened.
Site——Site source placed under version control — First committed snapshot of the hand-written site: home and docs pages, the self-serve pages, and nine model pages.
Cognitio1.0.0ReleasedFirst documented version. QLoRA adapter trained in job friend-lora-20260910-113709 (400 steps, 2,112 rows) on Qwen2.5-1.5B-Instruct; hosted endpoint serving since 2026-09-10.
Waymark1.0.0ReleasedFirst documented version. Served weights are the step-250 checkpoint (validation loss 0.1505); a refine run warm-started from it completed on 2026-09-09.
LIM1.0.0ReleasedFirst documented version. 512-wide single-bag classifier trained on Vertex AI (one L4); supersedes the 256-wide weights now listed as LIM Nano, with the same input and output contract.
LIM Nano1.0.0ReleasedFirst documented version. 256-wide single-bag classifier trained on a local CPU; predates LIM's 512-wide retrain and shares its contract. Serving on /v1/intention as lim-nano from 2026-09-13.
ELIM1.0.0ReleasedFirst documented version. Trained on Vertex AI (80,000 synthetic windows, 8 epochs, seed 7); serving as the default intention model since the same day.

Machine-readable: changelog.json