Hey everyone!
We just shipped TurboQuant into OpenCut AI, and this one changes what hardware you need to run the full AI stack.
The problem we had
OpenCut AI runs everything locally -- LLM, transcription, voice cloning, image generation. That's great for privacy, but brutal on memory. Running the full stack needed 35+ GB RAM. Most of our users have 8-16 GB laptops, so they were stuck with tiny 1B models that gave mediocre scripts, slow commands, and limited context.
What TurboQuant does
TurboQuant implements two algorithms from Google Research paper PolarQuant and QJL. That compress the KV cache (the biggest memory bottleneck during AI inference) by up to 6x with mathematically proven quality preservation.
In plain terms: your AI models now use a fraction of the memory without getting dumber.
Before vs After
On a 16 GB machine:
- Before: Llama 3.2 1B + Whisper Base + TTS = barely fits, mediocre quality
- After: Llama 3.1 8B + Whisper Medium + TTS = runs comfortably, dramatically better output
On an 8 GB machine:
- Before: Could only run the 1B model alone
- After: Runs a 3B model + Whisper Base + TTS together
Full stack memory:
- Before: 35 GB for everything
- After: 15 GB for everything
What this means for editing
- Better AI commands "remove the intro" actually works now because Mistral 7B understands context far better than a 1B model
- Better transcription Whisper Medium fits where only Whisper Base could before, so captions are more accurate
- Longer content: Process hour-long podcast transcripts without running out of memory. The 6x KV cache reduction means 6x longer input context
One-click setup in Settings
We added a new AI Optimization panel in Settings. It auto-detects your hardware and recommends the best configuration:
- Performance Tier: Lite (4-8 GB), Standard (8-16 GB), or Pro (16-32 GB). Each tier is tagged with "Best for your hardware" based on your actual RAM.
- KV Cache Compression: Pick 4-bit (near-lossless), 3-bit (5x compression), or 2-bit (aggressive). Recommended level highlighted based on your system.
- Memory Budget: Set once, and the system optimizes everything to fit.

Would love to hear, what's your RAM situation, and does this make local AI editing viable for you?

Hey ProductHunt!

We just shipped a big batch of features to OpenCut AI, our open-source self-hosted video editor. Here's what's new:

WebGL Transitions

OpenCut-AI just shipped real GPU support for TurboQuant KV cache compression.
OpenCut-AI is an open-source, local-first AI video editor. Everything runs on your machine transcription, voice cloning, image generation, LLM commands. No cloud, no API keys.
The catch was always memory. Running a 7B LLM + Whisper + TTS + Stable Diffusion locally means fighting for every gigabyte of RAM. TurboQuant solves this by compressing the KV cache (the biggest memory hog during inference) by up to 7.3 .
What's new in this release:
User-selectable Compute Mode in Settings AI Optimization. Pick Auto, CPU, or GPU (CUDA).
Real integration with the turboquant-gpu library. The GPU backend runs cuTile fused kernels for the full 2-bit / 3-bit KV compression path. The CPU backend uses a PyTorch fallback with physical-core thread pinning and MKLDNN acceleration.
Live-measured compression ratios in the UI. No more static lookup tables you see the actual compression your backend produced on the last request.
Graceful fallback everywhere. Missing CUDA? Falls back to CPU. Missing cuTile kernels? Falls back to PyTorch. The service always comes up.
Huge thanks to Anirudh Bharadwaj Vangara for the turboquant-gpu library that made the real GPU path possible.
OpenCut-AI: https://github.com/Ekaanth/OpenC...
turboquant-gpu: https://github.com/DevTechJr/tur...

Hey Hunters
We just shipped Google Gemma 4 support, paired with our TurboQuant KV-cache compression engine. That means you can now run Google's any-to-any multimodal models directly inside your editor no API keys, no cloud, no data leaving your machine.
What's new in this drop:
Full Gemma 4 family wired into the hardware-aware model registry:
- Gemma 4 E2B (5B) fits in ~3.5 GB, runs on 8 GB laptops
- Gemma 4 E4B (8B) ~5.5 GB, the new sweet-spot for Pro tier
- Gemma 4 26B MoE (4B active) big-model quality, efficient inference
- Gemma 4 31B Dense top-tier quality for 24 GB+ GPUs
TurboQuant KV-cache compression on every model:
- 3.8 compression at 4-bit (cosine similarity 0.9986 effectively lossless)
- 5.0 compression at 3-bit
- 7.3 compression at 2-bit for extreme memory savings
- Unlocks long-context editing sessions (32K 131K tokens) on consumer hardware
Hardware-aware auto-selection OpenCutAI detects your RAM/VRAM and picks the largest Gemma model that'll actually run smoothly. No guesswork.
Served through both Ollama (for simple local use) and our TurboQuant service
Why this matters:
Local video AI has always been a RAM problem. An 8B multimodal model + a long edit timeline + Whisper + TTS used to blow past 16 GB easily. With TurboQuant compressing the KV cache, you can now run Gemma 4 E4B end-to-end on a MacBook with room to spare.
Try it, tear it apart, tell us what breaks

Hey everyone!
We just shipped Smallest AI (Waves) integration into OpenCut-AI, and we're excited to share what this unlocks for creators.

What's new
OpenCut-AI now has three voice engines built in:
1. Local (Coqui XTTS) -- Runs entirely on your machine. 12 languages, voice cloning, zero API keys needed. Best for offline editing and privacy-first workflows.
2. Sarvam AI -- Purpose-built for Indian languages. 22 regional languages for transcription, 11 for text-to-speech, with 23+ natural speaker voices. If you're creating content in Hindi, Tamil, Telugu, Bengali, or any Indian language -- this is the best engine for you.
3. Smallest AI (NEW) -- Ultra-fast cloud TTS and STT. ~100ms latency, 80+ voices across 15 languages, and speech-to-text covering 39 languages with speaker diarization and emotion detection. This is our most versatile engine yet.
What you can do with Smallest AI
- Generate voiceovers in English, Hindi, Spanish, Tamil, and 11 more languages with natural-sounding voices
- Transcribe audio/video in 39 languages with automatic subtitle generation
- Control speech speed from 0.5x to 2.0x
- Pick from 80+ voices -- each language has multiple male and female options
- Process long content -- auto-chunking handles videos of any length
How it works
1. Grab a free API key from app.smallest.ai
2. Paste it in Settings > API Keys > Smallest AI
3. Select "Smallest AI" in the Voiceover or Captions panel
4. Generate
That's it. No server setup, no model downloads, no GPU required.
We're open to adding more models
This is the part we're most excited about. OpenCut-AI is built with a pluggable engine architecture. Adding a new voice or transcription provider is straightforward, and we want the community to drive what comes next.
Models we're considering:
- ElevenLabs (premium voice synthesis)
- Deepgram (real-time STT)
- Fish Speech (open-source voice cloning)
- Kokoro TTS (lightweight and fast)
- StyleTTS 2 (human-level quality)
Have a model you'd love to see? Drop it in the comments with:
- What it does
- Why it matters for video editing
- A link to their docs
We'll prioritize based on what the community wants most.
Links
- GitHub: github.com/Ekaanth/OpenCut-AI
- Smallest AI Docs: waves-docs.smallest.ai
- Get a Smallest AI key: app.smallest.ai
We'd love to hear what models and features you want next. Let us know!