Models
Every AI model in OpenCut AI is open source and runs locally on your machine. Cloud APIs available for Indian and multilingual content.
Performance Tiers
Lite
4-8GB
Llama 3.2 1B + Whisper Base + noisereduce
Runs on any machine. Basic AI commands, transcription, and audio cleanup. TurboQuant lets you squeeze a 3B model in here.
Standard
8-16GB
Llama 3.2 3B / Mistral 7B + Whisper Small + XTTS v2 + rembg
Recommended for most users. Good AI commands, voice cloning, and image generation. TurboQuant compression makes this the sweet spot.
Pro
16-32+GB
Llama 3.1 8B + Whisper Medium + SDXL + XTTS v2 + full stack
Best quality across the board. 8B models with near-lossless 4-bit compression. GPU recommended for image generation.
TurboQuant Memory Compression
Built-inAll tiers benefit from TurboQuant KV cache compression. Full AI stack memory dropped from 35 GB to 15 GB. One-click setup in Settings auto-detects your hardware and recommends the best configuration.
4-bit
0.9986 cosine sim
Near-lossless
3-bit
0.9953 cosine sim
6x compression
2-bit
0.9874 cosine sim
7.3x compression
All models
Whisper
OpenAI
Automatic speech recognition with word-level timestamps. Powers transcription, subtitles, and text-based editing. Available from tiny (39M) to large-v3 (1.5B). TurboQuant compression lets Whisper Medium fit where only Base could before.
Llama 3.2 / 3.1
Meta
Powers AI commands, script writing, chapter analysis, clip finding, smart suggestions, and prompt enhancement. Runs via Ollama. Available at 1B (Lite), 3B (Standard), and 8B (Pro) with Q3/Q4/Q5 quantization.
Mistral 7B
Mistral AI
High quality 7B instruction-tuned model. Available at Q4 (Standard tier) and Q5 (Pro tier). Strong at structured JSON generation for editor commands and content analysis.
Qwen2.5
Alibaba
TurboQuant-validated model family from 0.5B to 14B. The 3B Instruct variant is benchmarked with full compression data (0.9986 cosine similarity at 4-bit). Also includes Coder variants for technical content.
Stable Diffusion XL
Stability AI
Generate images from text prompts for video overlays, thumbnails, and B-roll. Supports SDXL Turbo (4-step fast generation) and full SDXL (20-step quality).
FLUX.1
Black Forest Labs
Next-generation text-to-image with superior prompt adherence and image quality. Alternative to SDXL for higher quality results.
XTTS v2
Coqui
Local voice synthesis with voice cloning from a 6-second sample. Supports 17+ languages. TurboQuant at 3-bit shrinks it from 1.8 GB to 0.6 GB while preserving voice quality.
Sarvam AI
Sarvam AI
Purpose-built for 22 Indian regional languages. Saaras v3 for transcription, Bulbul v3 for text-to-speech with 37+ natural voices, plus translation and transliteration.
Smallest AI
Smallest AI
Ultra-fast cloud TTS and STT. Lightning TTS with 80+ voices across 15 languages at ~100ms latency. Pulse STT covers 39 languages with speaker diarization and emotion detection.
Pyannote
Herve Bredin
Neural speaker diarization that identifies who is talking when. Powers multi-speaker detection and automatic speaker-boundary cuts. Falls back to FFmpeg silence-based detection.
MediaPipe
Real-time face detection for auto-reframe. Tracks faces across video frames to generate 9:16 crops for TikTok, Reels, and Shorts.
SpeechBrain
SpeechBrain
Detects emotional peaks in audio for finding the most impactful moments. Used for clip scoring and highlight detection. Falls back to energy-based analysis locally.
TurboQuant
OpenCut AI
KV cache compression using Google Research's PolarQuant + QJL. Achieves 6x memory reduction at 3-bit with 0.9953 cosine similarity. Makes 7B models run on 8 GB RAM. One-click setup in Settings.
U-Net (rembg)
danielgatis
Remove backgrounds from images to create transparent overlays. Useful for speaker cutouts, product shots, and compositing.
noisereduce
Tim Sainburg
Spectral gating noise reduction for cleaning up audio recordings. Removes background noise, hiss, and hum with adjustable strength.