Powered by SharpMind
SharpMind Studio · 1.1.0

Your models.
Your machine. Or a GPU you rent for an hour.

One workbench for chatting, building datasets, fine-tuning with LoRA, and running models. Start on this PC. Add a SharpMind Compute host on Runpod or Vultr when the job needs more memory — and keep the same workflow, token and data layout.

Download for Windows Set up a GPU host → Read the wiki
Local · your weights LoRA · GPU fine-tuning Runpod · API deploy Vultr · SSH setup HuggingFace · GGUF Pairing · phone/LAN OpenRouter · remote models
Windows installer on GitHub Releases. The repository is the source of truth; the wiki tracks setup, GPU hosts and troubleshooting.
SharpMind
Fine-tune Qwen on a rented A16.
SSH setup done. Pulling the GPU image…ports 5117 + 8080
And the dataset?
Uploaded with the job. Export comes back here.merged .smm or .smlora
● GPU Runpod Vultr
Message the model…
The brief

A local-first workbench that can borrow a GPU.

SharpMind Studio turns a folder of text into a model you can chat with, then gives that model projects, tools and a workspace. Everything stays on your PC until you explicitly send a job to a host or a prompt to OpenRouter.

01

Build

Data Factory, pretraining, LoRA fine-tuning and SFT — with checkpoints, validation and a promotion exam.

02

Run

Local chat and Work, OpenRouter, or your own Compute host with an OpenAI-style API and GPU inference.

03

Keep

Models, sessions, checkpoints and exports stay yours. Encrypted keys, revocable phone pairing.

One workbench

Chat, data, training and inference without the glue code.

SharpMind Studio is the front end for the SharpMind libraries. Model formats, tokenizers, checkpoints, chats and run records live in one place — on your PC, or on a host you control.

01

Train, on the right device

From-scratch pretraining stays on the CPU. LoRA fine-tuning and instruction tuning run on an NVIDIA GPU when the base model fits the engine.

pretrain · LoRA · SFT · CPU fallback
02

Compute hosts, not terminals

Local Docker, Runpod through its API, and Vultr or any Linux box through the built-in SSH setup. Start, test, log and stop from Settings.

Local · Runpod · Vultr · custom
03

Data Factory

Source → clean/filter → provider transform → review → export as chat or text JSONL, ready for SFT or pretraining.

saved .smmdf recipes · previews
04

Models and formats

Browse HuggingFace GGUF repositories, download with resume, convert to .smm, and package exports for inference.

.smm ↔ .gguf · .smlora
05

Chat and OpenAI-style API

A Compute host serves one model per process with streaming chat, uploads, model switching and background training jobs.

/v1/chat/completions · SSE · bearer token
06

Work mode

Project context, scoped file and web tools, skills, and an explicit permission gate — local tools still apply with a remote model.

project memory · permissions
A look inside

Chat, tools and token speed in one view.

A local Qwen2 0.5B model answering a prompt. The header shows the model, the tool probe and the measured throughput — no cloud round trip.

SharpMind Studio chat view: a local Qwen2 0.5B model answers a story prompt, with tools ok and 14.10 tokens per second in the header.
Version 1.1.0 · local Qwen2 0.5B instruct, chat view. Models & chat guide →
GPU fine-tuning

LoRA on a rented GPU.
The model stays portable.

The GPU engine trains adapters, not whole weights. Upload a packaged .smm base with the job, train on the host, then download a merged model or an adapter back to this PC. No data leaves except the job you explicitly send.

  • LLaMA/Qwen-style base: RMSNorm, RoPE or NoPE, gated FFN.
  • From-scratch pretraining and MoE bases stay on the CPU.
  • An unsupported base is refused with a reason, then retried on the CPU.
  • Inference also covers LayerNorm and dense-GELU models.
CUDA · cuBLAS ILGPU OpenCL merged .smm adapter .smlora 
# on the host, /health is honest about it
{
  "status": "ok",
  "accelerators": [{
    "name": "ilgpu",
    "inference": true,
    "backend": "CUDA · cuBLAS"
  }]
}

# a fine-tune log that used the device
Fine-tune: 8 adapters (rank 8, alpha 16).
Engine: [Cuda] NVIDIA A16, 14962 MB, cuBLAS 12.6
Training 100 steps (float32)…

# if the base is unsupported:
Accelerator 'ilgpu' declined: LayerNorm, only RMSNorm
Retrying on the CPU. Set SHARPMIND_ACCELERATOR_FALLBACK=false
to fail instead.
Four places to run

Choose the host. Keep the workflow.

Every host runs the same image contract: one model, ports 8080 and 5117, and a shared /data/SharpMind folder.

LOCAL

On this PC

Full privacy and zero token cost. Run a packaged model or convert a GGUF. Pretraining works on the CPU.

LOCAL DOCKER

The Compute API at home

Studio starts and stops the container with the saved token, port and folder — the compute image, or the GPU server image.

RUNPOD

GPU by the minute

Studio creates the pod through the Runpod API, wires the proxy URLs and can terminate it. Account balance is required.

VULTR

A server you keep

Create the instance, let Studio log in over SSH, install Docker and the NVIDIA toolkit, pull and run the container.

Guides

Start with the host you have.

The wiki has the full setup path, the exact error messages, and what to do next.

GPU · SSH

Vultr + SSH setup

Create an A16, add a key once, then let Studio install Docker and deploy the GPU image.

~15 minutes · billed hourly Local

Local Docker

Build the combined server image on this PC and start it from the Compute card.

CPU or local NVIDIA GPU Cloud

Runpod

Deploy a pod through the API, expose both ports, and terminate it when the run is done.

API key · Docker Hub · proxy URLs
Compare providers How training works Troubleshooting
Build from source

1.1.0, from source.

The repository is the source of truth. Build, test and run the Studio locally; publish your own server image when you are ready to use a host.

No packaged installer yet. NuGet packages and the public repository are the release path; the wiki tracks both. 
# clone and build
git clone https://github.com/Integral2u/SharpMindStudio
cd SharpMindStudio
dotnet build SharpMind.Studio.slnx -c Debug
dotnet test  SharpMind.Studio.slnx -c Debug
dotnet run --project SharpMind.Studio

# open the local Studio
http://127.0.0.1:5117

# add a host under Settings → Providers → Compute