Just want to see it run?
Build the app, pick a GGUF or .smm, and chat locally. No account and no host.
The same Studio, the same model format, four places to run it. Start with the machine in front of you; add a GPU host when a fine-tune needs more memory. Every guide includes the errors you are likely to see and what they mean.
A quick tour of the local-first workbench: models, chat, Work, training and Labs. The playlist carries the longer walkthroughs.
GPU · SSHCreate the instance, add one key, and let Studio install Docker, the NVIDIA toolkit and the server image.
LocalBuild the Compute image on this PC (or the GPU server image) and start it from Settings.
CloudDeploy a GPU pod through the Runpod API and terminate it when the run is done.
ModelsBrowse GGUF repositories, check licence and size, download with resume and convert to .smm.
Remote modelsConnect hosted models for Chat and Work with an encrypted key and local tool permissions.
ProjectsProject context, scoped file and web tools, skills and explicit read/write/delete permissions.
LabsLayer importance, quantization, activation maps, latent space, architecture templates and agent definitions.
ConfigurationAppearance, system tools and plugins, family and model defaults, providers and pairing.
AccessOpen the local Studio from a phone on your LAN with a QR or PIN, then revoke the device when you are done.
OverviewLocal Docker, Runpod, Vultr or your own Linux box: cost, ports and who runs the setup.
TrainWhich jobs use the GPU, which stay on the CPU, and why a LayerNorm base falls back.
ErrorsPull access denied, accelerator declined, firewall timeouts, SSH auth, 401 after a token rotation.
This PC covers the Studio process and an optional Local Docker host on the same machine. A local NVIDIA GPU is used only when the GPU image and Docker runtime are present. A Compute host always brings its own device.
| Capability | This PC / Local Docker | Compute host | Notes |
|---|---|---|---|
| Chat with a packaged model | Yes | Yes | One active model per process on a host. |
| OpenAI-style API | — | Yes | Bearer token; /v1/chat/completions, models, files, training jobs. |
| Pretrain from scratch | Yes (CPU) | Yes (CPU) | The GPU engine refuses whole-weight training by design. |
| Layer importance lab | Yes | - | Model-agnostic per-block attention/FFN ablation with activation RMS and loss deltas; source builds. |
| LFM3-VL image input | Yes | - | Attach an image in Chat when the vision mmproj is next to the model; the native tower and 2x2 merge projector describe it. Source builds. |
| LoRA fine-tune | Yes (GPU or CPU) | Yes (GPU or CPU) | GPU needs RMSNorm + RoPE/NoPE + gated FFN; otherwise the CPU retry runs. |
| SFT (assistant-only loss) | Yes (GPU or CPU) | Yes (GPU or CPU) | Chat JSONL; the base supplies the tokenizer and template. |
| GPU inference | With a supported local GPU | Yes (GPU or CPU) | Also covers LayerNorm and dense-GELU models. |
| Local chat artifacts | Send / render / mic | Not yet | Images, audio, files and code. LFM3-Audio microphone and audio attachments transcribe locally. |
| LFM3 text (13 conv / 5 SWA / 6 global) | Smoke | Smoke | Layer map, per-head gate and templated chat validated; broad quality pending. |
| ASR / TTS | ASR and TTS | No | Native LFM3-Audio encoder, MaskGIT depth decoder and LiquidCodec via mmproj; long recordings are chunked and speech is written as a 24 kHz WAV. |
| Remote data and artifacts | — | Yes | Uploads with progress; exports download back to Studio. |
Build the app, pick a GGUF or .smm, and chat locally. No account and no host.
Rent an A16, add one SSH key, and let the Compute card install and deploy the GPU image.