GhostSwitch
ExperimentSmall model. Expert answers.
An AI assistant that runs entirely on your computer. It keeps one Qwen3.5 model in memory and, for each message, attaches the LoRA expert that fits: code, finance, law, math, Verilog, CAD and two dozen more. Then plain code checks the answer before you see it.
How it works
One base model stays loaded. A 30–330 MB adapter is attached only for the request that needs it, so a library of one expert or a thousand needs the same memory: the base model plus at most 512 MB of recently used adapters.
Pick the expert
Short model calls walk a tree of fields and choose among the experts you turned on, or the plain base model.
Swap the adapter
The LoRA goes onto the model already in memory in 0.1–0.3 s. No reload.
Answer
Optionally with web search and cited sources, and step-by-step thinking when the task needs it.
Verify in code
Arithmetic, JSON fields and citations are checked by code, not by another model. A failed check means a retry.
Measured, not assumed
An adapter that sounds good on its model card does not always beat the model it was trained on. 13 of the 9B experts each got 60 questions from a public benchmark of their field, and the bare base model got the same 60. Experts that are no better are blocked from routing.
| Expert | Benchmark | Base | Expert | p | Verdict |
|---|---|---|---|---|---|
| coding | HumanEval+ | 76.7% | 88.3% | 0.039 | Verified |
| legal contracts | LegalBench CUAD | 76.7% | 83.3% | 0.125 | Leaning better |
| privacy / PII | Gretel PII, F1 | 75.3% | 78.3% | 0.180 | Leaning better |
| software engineering | HumanEval+ | 76.7% | 81.7% | 0.508 | Undecided |
| finance | FinQA | 61.7% | 66.7% | 0.549 | Undecided |
| reasoning | GSM8K, no thinking | 88.3% | 90.0% | 1.000 | Undecided |
| DeltaCoder | HumanEval+ | 76.7% | 78.3% | 1.000 | Blocked |
| text-to-SQL | Gretel, run in SQLite | 58.3% | 58.3% | 1.000 | Blocked |
| medical | MedQA | 66.7% | 65.0% | 1.000 | Blocked |
| pentest | pentesting-eval | 88.3% | 86.7% | 1.000 | Blocked |
| IaC security | Terraform S3 MCQ | 53.3% | 51.7% | 1.000 | Blocked |
| entity extraction | CoNLL-2003, F1 | 67.4% | 67.2% | 0.720 | Blocked |
| CWE classification | CTI-Bench | 56.7% | 51.7% | 0.508 | Blocked |
Qwen3.5-9B Q4_K_M, answer checks and thinking off for both sides, paired sign test. 60 questions detect differences of about 10 points, so “undecided” means undecided. Every answer is in the evaluation dataset.
Install
One command installs GhostSwitch and opens it in your browser. It brings its own Python and needs no administrator rights. On first run it checks your hardware, recommends Qwen3.5-9B or 4B and downloads it.
curl -fsSL https://raw.githubusercontent.com/hiddenway/GhostSwitch/main/install.sh | shirm https://raw.githubusercontent.com/hiddenway/GhostSwitch/main/install.ps1 | iex16 GB of memory
Qwen3.5-9B with 27 experts. 8 GB is enough for the 4B model with 6 experts.
MLX or llama.cpp
MLX on Apple Silicon; llama.cpp with CUDA on NVIDIA, or on the CPU, on Linux and Windows.
OpenAI-compatible API
Streaming, function tools, images and files at /v1/chat/completions. Works as a model in OpenCode.
Private by default
Chats stay on your computer. It goes online only to download models and experts, and for web search if you turn it on.