Transcription
Meetings and lectures become timestamped text.
- Handles hours-long recordings with smart chunking
- Markdown, JSON, SRT and VTT export
- Live progress, stoppable anytime
Private AI workbench for your DGX Spark
One box. Six models. Zero cloud required.
Transcribe meetings. OCR documents. Translate, chat, generate images, and build mind maps. Then pass any result straight into the next tool, or chain them into visual workflows. Everything stays on your machine, in one folder you can simply copy to back up.
Install SparklingKit
Docker required
curl -fsSL https://run.sparklingkit.com/stable/install.sh -o install.sh && bash install.sh
Install the workbench wherever you want itRun this on your DGX Spark or another machine; guided model setup comes next.
Then open
http://localhost:54321The thesis
A gaming GPU forces you to pick one model. 128 GB doesn't.
Unified-memory machines changed the rules of local AI. With 128 GB, you don't choose between a chat model or a transcription model or an image model. You keep a whole team of specialists loaded at once, each one excellent at its job.
SparklingKit is the workbench that puts that team to work: one interface, one folder, one small box on your desk that quietly does what used to take five cloud subscriptions.
Six tools, one panel
Everything you'd reach for, already there.
OCR
Scans and PDFs become clean Markdown.
- Batch multiple files
- Structured page data
- Originals preserved beside every result
Translation
A dual-pane translator, powered by a dedicated model.
- Live translation as you type
- Automatic language detection
- Works with pasted text or processed documents
Visual Grounding
Ask an image where things are.
- Up to 12 natural-language queries per job
- Boxes drawn directly on the image
- Coordinates available downstream
Text to Image
Generate images locally. No credits, no queue.
- Square, landscape and portrait presets
- Prompt saved with every image
- Full thumbnail history
Mind Maps
Any topic, note, or document becomes an interactive map.
- Pan, zoom, expand and collapse
- Configurable depth and breadth
- Portable JSON plus a Markdown outline
Chat
Talk to your LLM or to any result above.
The Continue with system
Your work doesn't end. It continues.
In most AI tools, a result is a dead end. You copy, paste, and start over somewhere else. In SparklingKit, every result is a typed artifact that knows where it can go next.
Finish a transcription and the next steps are already waiting: Translate it. Map it. Chat with it. OCR a contract, then question it. Generate an image, then ground objects inside it. Each tool declares what it accepts and produces, and SparklingKit routes the possibilities automatically, including into any workflow you've enabled.
One upload, an entire chain of work, all without leaving the panel.
meeting-transcript.mdTranscription · completed just now
TranslateAccepts transcript
Mind mapAccepts document
ChatAttach as reference
Visual workflows
Chain it once. Run it forever.
The same tools you use by hand snap together in a node-based workflow editor. Drag in OCR, Translation, an LLM prompt, a Mind Map. Connect them. Incompatible connections are simply refused, so broken graphs can't happen. Add conditions, branches, and merges when tasks get smarter.
Then watch it run. The live graph shows every node working, finishing, or being skipped, and every past run keeps its full diagram: which branches fired, how long each step took, and the complete lineage from input to output.
- Workflows are saved as readable JSON files.
- Enabled workflows appear on the Workbench and in Continue-with menus.
- No arbitrary code execution. Workflows are data, not scripts.
Lecture to study kitRunning
Inputlecture.pdf
OCR8 pages
TranslateProcessing
Mind mapWaiting
Revision 4 · JSON workflowEvery node keeps its status and lineage
Service nodes
- OCR
- Transcription
- Translation
- Grounding
- Text to Image
- Mind Map
- LLM Prompt
- Create Chat
Logic nodes
- Input
- Select
- If
- Switch
- Merge
- Save to File
- End
- Fail
Everything is a file
Your workspace is a folder. That's the whole story.
Every job, transcript, image, mind map, chat, workflow, and setting lives as an ordinary file inside one directory. No database server. No proprietary formats. No export button because there's nothing to export from.
Back up by copying the folder. Migrate by moving it. Search it with your own tools. Point Obsidian at it.
Preview, rename, download, or delete anything from the panel, or just open the folder. It's your data, in the most literal sense possible.
sparklingkit/data
config/settings and workflows
settings.json
workflows/
lecture-to-study-kit.json
jobs/sources and outputs
meeting-transcription/
meeting.m4a
transcript.md
mindmap.json
chats/referenced conversations
project-review/chat.json
logs/
The hand-picked stack
Six models, chosen to share one machine.
We didn't grab the biggest model of each kind. We picked specialists that are excellent at their jobs and fit together in 128 GB with room to breathe. That's the unified-memory advantage, used deliberately.
Multimodal LLMQwen3.6-35B-A3B (NVFP4)Chat, reasoning, vision and workflow prompts
OCRUnlimited-OCRPage and image text extraction
Speech recognitionQwen3-ASR-1.7BAudio and video transcription
TranslationHy-MT2-1.8B (FP8)Dedicated multilingual translation
Visual groundingLocateAnything-3B*Query-driven image region location
Image generationZ-Image-TurboFast local text-to-image generation
The starter tooling handles everything between bare machine and working stack: architecture checks, model downloads with pinned revisions, container builds, memory-conscious startup order, and per-service readiness checks.
And none of it is hard-coded. Every model is independently replaceable. SparklingKit is the workbench and orchestration layer, not a locked bundle. Swap any endpoint for your own favorite.
Model weights are downloaded from their original publishers and each retains its own license. * LocateAnything-3B is currently published under research/non-commercial terms. Swap the grounding endpoint if your use is commercial.
Run it your way
From one box to any topology.
All-in-one
SparklingKit and all six models on a single DGX Spark. The full workbench in one small machine on your desk.
Split
Models on the Spark, SparklingKit on the server that owns your data. Your workspace lives where you want it.
Bring your own
Just the panel, pointed at compatible local servers, remote machines, cloud APIs, or a deliberate mix. Local-first, your call always.
First launch
Onboarding that respects your time.
Choose models on this machine, connect to a remote machine, or configure endpoints by hand. Each path includes copyable commands, a verification check for every service, and a staged review before anything is applied.
Run locallyRun remotelyConfigure manually
Multimodal LLM Verified
OCR and transcription Verified
Translation and grounding Verified
Image generation Verified
Live monitoring
A machine you can see.
Every service shows live health in the sidebar. Add the optional DGX Spark monitor for unified-memory use, CUDA allocations, GPU utilization and temperature at a glance.
DGX SparkLive
VRAM101.7 GiB / 121.7 GiB
CUDA allocations82.6 GiB
GPU 18% · 48°C6/6 models
Updates without fear: the built-in updater verifies checksums, keeps your data folder untouched, waits for health checks, and can roll back to the previous version with one command.
Private by architecture
Nothing leaves unless you send it.
SparklingKit runs on your hardware and stores everything in your folder. No accounts. No telemetry. No cloud dependency. Meetings, contracts, and recordings are processed on the machine under your desk and stay there.
Local-first, not local-only: connecting a cloud API is always your explicit choice, never a default.
Atomic writesJob records are protected from partial updates and corruption.
Separated secretsAPI keys live apart from job data with restrictive permissions.
Durable recoveryInterrupted work reconciles from durable files, not fragile queues.
Readable foreverEvery record remains an ordinary file in formats you already know.
SparklingKit doesn't include multi-user authentication yet. For access beyond your own network, place it behind a VPN or authenticated reverse proxy.
FAQ
The practical questions.
Do I need a DGX Spark?
For the complete out-of-the-box model stack, currently yes. SparklingKit is tested and supported on NVIDIA DGX Spark with the GB10 Grace Blackwell Superchip. We are adding support for other unified-memory machines. You can already run the SparklingKit panel separately and connect it to compatible local, remote, or cloud endpoints.
Are the models included in the app?
No. The app image stays small. The starter tooling downloads pinned model revisions from their original publishers on first setup, and reuses downloads and caches afterward.
Is it really free?
SparklingKit is Apache 2.0. Each model keeps its own license. Most are permissive; check the model-stack footnote for the current exception.
What happens to my data if I stop using SparklingKit?
Nothing. It is already yours, as Markdown, JSON, images, and media files in one folder. There is no lock-in to escape from.
Can I use it on my phone?
The panel is fully responsive. Start a transcription from your phone on the LAN and check progress from anywhere in the house.
Can workflows run untrusted code?
No. Workflow definitions are pure data and cannot execute JavaScript or shell commands.
Make the hardware yours
One small box. A complete AI workbench. Yours.
Install SparklingKit
Docker required
curl -fsSL https://run.sparklingkit.com/stable/install.sh -o install.sh && bash install.sh
Install the workbench wherever you want itRun this on your DGX Spark or another machine; guided model setup comes next.
Then open
http://localhost:54321