How-to
Help articles for every feature
Each card in the Blueprint desktop app links here. Articles cover the background you need to understand what a feature does, plus how to use it. Same content; in-app and on the web both render from the same source.
Plan
Pick a model, size it, see what it would cost.
Picking a model
How to choose an open LLM for your hardware and workload — the planner walks you through it; here's the background it assumes.
2 min · Plan
VRAM math
How Blueprint estimates GPU memory for a (model, quant, context, batch) combination — and why the number on the screen is the number, not a guess.
2 min · Plan
Hardware
VRAM math, tier recommendations, on-prem vs cloud.
On-prem vs cloud
The break-even point between buying GPUs and renting them comes faster than most teams expect — Blueprint shows you exactly where.
3 min · Hardware
Sizing hardware
Read the tier recommendations, understand why each one's labeled what it is, and pick the box that fits your workload without overpaying.
2 min · Hardware
Deploy
Install runtime, pull GGUFs, supervise llama-server.
Installing the runtime
The runtime is llama.cpp — Blueprint downloads the pre-built binary for your OS and verifies it before letting you serve anything.
2 min · Deploy
Pulling models
How Blueprint downloads GGUF weights from HuggingFace, where they go, and how to resume an interrupted pull.
3 min · Deploy
Starting llama-server
Once the runtime is installed and a model is on disk, the Serve action spawns llama-server with the right flags — here's what those flags do.
3 min · Deploy
Verifying a deploy
After Start serving, the Verify chat is your smoke test — send a prompt, watch the response, confirm the model is actually doing what you expected.
2 min · Deploy
Dashboard
Service, router, prompt cache, Python runtime, engines.
Choosing an engine
llama.cpp vs vLLM vs TensorRT-LLM — when each one wins, why Blueprint defaults to llama.cpp, and when to switch.
3 min · Dashboard
Inference chat
The Dashboard's Inference tab is a tuned chat UI against whatever's currently serving — local, remote host, or registered remote server. Sampling controls are exposed so you can dial in for your workload.
3 min · Dashboard
Install Blueprint as a system service
Run blueprint-svc as a Windows Service or systemd unit so models stay served across reboots, without keeping the desktop GUI open.
3 min · Dashboard
llama-server metrics
The Performance card reads llama.cpp's Prometheus metrics endpoint — here's what each number means and which ones to actually optimize for.
3 min · Dashboard
LoRA training
Fine-tune a low-rank adapter on your data with QLoRA — runs on a single 24 GB consumer GPU, outputs a 50 MB adapter that loads cleanly into llama.cpp.
3 min · Dashboard
Models on disk
What the Dashboard's Models card shows — installed GGUF files, where they live, how much disk they use, and how to pull more.
3 min · Dashboard
Prompt cache
Cache the KV-state of common prompt prefixes so the next request with the same prefix skips the prefill phase entirely.
3 min · Dashboard
Python runtime
The optional Python sidecar Blueprint uses for the heavy bits — LoRA training, LLMLingua compression, vLLM, TensorRT-LLM. Opt-in feature by feature.
3 min · Dashboard
Remote OpenAI-compatible servers
Register an LLM endpoint someone else operates — your shared team server, a colleague's box, a vLLM instance behind nginx — and reach it from the Dashboard.
3 min · Dashboard
Semantic router
Route easy queries to a fast small model and hard queries to your big one — drops average cost per query by an order of magnitude without dropping the ceiling on hard ones.
3 min · Dashboard
Calibrate
Custom imatrix, eval, calibrated quantization.
Calibrated quantization
Step 4 of the Calibrate workflow — turn your imatrix into actual GGUF files for the target quant levels you want to ship.
2 min · Calibrate
Custom imatrix calibration
Calibrate a quantization against your own prompts so the bits land where YOUR workload pushes activation magnitude — typically 8-15% accuracy gain at the same file size.
3 min · Calibrate
Evaluating quants
Step 5 of the Calibrate workflow runs every candidate quant against your held-out eval set and gives you the numbers to compare — quality, latency, throughput per candidate.
3 min · Calibrate
Hosts
SSH-managed remote machines and fleet operations.
Adding a host
Register a Linux GPU box so Blueprint can SSH in, push-install the svc, and drive llama-server on it from one desktop.
3 min · Hosts
Push install
Install blueprint-svc on a registered remote host without ever SSHing in yourself — Blueprint ships the binary over the existing SSH connection and registers it as a systemd unit.
3 min · Hosts
SSH tunneling
How Blueprint reaches a remote host's svc over the existing SSH connection — no port forwarding, no firewall changes, no `ssh -L`.
3 min · Hosts
Maintain
Update runtime, swap models, restart, clean up.
Resetting Blueprint data
How to clean up models, calibration runs, the Python sidecar, or the entire ~/.blueprint directory — including what survives a reset and what doesn't.
3 min · Maintain
Updating the runtime
When llama.cpp ships a new release with a feature or fix you want, the Maintain tab pulls the latest build and swaps it in — without disturbing your running models.
2 min · Maintain