NVIDIA's multilingual Nemotron ASR Streaming model, matching the 600M-parameter Nemotron 3.5 ASR checkpoint from June 2026: cache-aware streaming transcription across 40 language-locales (32 usable out of the box) with punctuation and automatic language detection.
FLUX.2 [dev], Black Forest Labs' 32B open-weight image model (November 2025) with a Mistral 3 vision-language encoder, better text rendering and output up to 4MP. This endpoint is text-to-image only; BFL's FLUX 3 family (July 2026) is newer.
FLUX.1 Krea [dev], Black Forest Labs' July 2025 open-weight text-to-image model built with Krea AI to avoid the oversaturated 'AI look'. It shares FLUX.1 [dev]'s 12B architecture; FLUX.2 and FLUX 3 have since superseded it.
FLUX.1 [dev], the 12B guidance-distilled text-to-image model from Black Forest Labs' August 2024 launch, released under a non-commercial weights license. It is two generations old: FLUX.2 [dev] (November 2025) and FLUX 3 (July 2026) followed.
FLUX.1 [schnell], the Apache 2.0, timestep-distilled member of Black Forest Labs' first FLUX family (August 2024), generating images in 1 to 4 steps, trading fidelity for speed. FLUX.2 [klein] is BFL's newer small open model.
OpenAI's Whisper large-v3 (November 2023) with options the bare model lacks: speaker diarization, word- or segment-level chunks, and translation of about 99 languages into English. OpenAI's later large-v3-turbo trades some accuracy for speed.
Wizper is fal's experimental inference-optimized build of Whisper v3 Large, which fal says matches the original's word error rate at about twice the speed. It supports transcription and English translation with timestamps, but not speaker diarization.
fal's real-time transcription endpoint built on NVIDIA's Canary speech recognition models; it streams partial and final transcripts but exposes only a punctuation toggle, with no language, timestamp or translation options. fal does not name the Canary checkpoint.
Queue-based 'turbo' variant of fal's Canary-powered transcription endpoint, priced the same per second of audio and taking only an audio URL plus a punctuation toggle. fal publishes no speed figures or checkpoint name.
