Choose models for your computer and purposes, or build a RAG system. Compare evidence, compatibility and estimated memory before downloading.
—catalogue computers
—local model choices
0models in your plan
Current plan
Choose an intended use to begin.
YOUR WORKBENCH
A setup, shaped around you.
▣
Selected in your RAG plan
Browse models
These are available choices, not selected models. Choose a stage, then use its selection button to add a model.
RAG pipeline stages
Nine-stage RAG interface
Choose a stage to browse candidates or baseline script settings. The planner describes a proposed pipeline; it does not execute Python scripts.
Your roles. Only as many as you need.
Start with one model or build the full nine-role architecture. Matching selections share one serialized instance.
RAG planning mode
Custom RAG baseline
Adjustable estimates for a document retrieval pipeline. No files are uploaded or read.
Advanced RAG settings
Format changes apply to eligible generative models. GPT-OSS stays MXFP4; CPU encoders keep their own format. Cache precision is not changed.
BEFORE YOUR FIRST DOWNLOAD
Know what you’re bringing home.
Purpose evidence checked 25 October 2026
01 / THE FILE
A model name isn’t a file format.
GGUF, MLX and Safetensors
GGUF is the route used by this planner for generative models, with llama.cpp through Vulkan, Metal, CUDA or CPU. Start with Q4_K_M for eligible models. Q5/Q8 cost more disk and RAM; quality improvements depend on the task. GPT-OSS uses its own MXFP4 representation.
MLX is an Apple Silicon alternative. Use an MLX-compatible converted model and a runtime that supports that exact architecture. GGUF and MLX artifacts are not interchangeable. Safetensors is a storage format used by Transformers and MLX; the model configuration and runtime still have to match.
Use the official model card, then inspect its linked quantizations. Check uploader, exact family/size, Instruct or IT variant, quantization and any vision projector. A source link here is not an automatic download.
02 / THE WORKFLOW
Get one thing working first.
Windows, Mac and Linux routes
LM Studio offers a convenient desktop path on supported Windows and Apple Silicon systems. Use Vulkan for the AMD baseline; Metal on Mac; CUDA for a supported NVIDIA GPU. ROCm support depends on the exact hardware, OS, driver and framework combination.
GB10 devices use Linux/Arm64. Follow the NVIDIA software guidance and an appropriate CUDA build. Do not download x86 Windows tools for an Arm Linux machine.
For a local RAG application, use a local endpoint, Python, a single-process Qdrant local store, and SQLite metadata/BM25. CPU encoders use their own Transformers/Sentence Transformers path. Bind to localhost and disable cloud fallback if the intent is fully local.
03 / THE DECISION
Fit is the beginning.
Try your own ten representative prompts. Compare answer quality, time to first response, memory pressure and mistakes. A larger model or a higher TOPS figure does not automatically mean a better everyday assistant.
Origins, licenses & limitations
Model origin means the producer location represented in the catalogue. It is not a claim about hosting location, data-storage location, every training-data source, every contributor, or every quantization uploader.
Licence labels summarize the cited model metadata. Permissive licences can still require product, legal and security review for a specific use. Custom or restricted terms are not all the same; inspect the exact model card or provider terms before using a model commercially.
Product names belong to their respective makers. Device cards link manufacturer specifications; announced products are marked. Prices, local availability, warranties, shipping SKUs and delivery dates must be checked before purchase. Bandwidth is a published specification, not a measured generation speed.
This site is a planner. It does not install software, download model weights, run inference, or read your files. Settings stay in this browser.