Local coding
Lokales Coding-Modell aufbauen
Unlimited supply. Absolute privacy. Built for heavy users. Tokens that never meter out, repos that never leave your machine—pick stack, model, and hardware in that order (or reverse from the GPU you already own).
Recommended order
-
1. Software stack
Ollama to start, llama.cpp for control, MLX on Apple Silicon, vLLM/SGLang only on a dedicated machine. Do not begin with vLLM on a laptop.
Software-Stack → -
2. Coding model
Only downloadable open weights that fit a consumer band. Site global rank is not a SWE score. Skip 480B-class MoE for a desk GPU.
Modelle → -
3. Hardware
VRAM (or unified memory) is the lock—not TFLOPS. Weights fitting the card ≠ your target context fitting. Pay electricity and depreciation, not token stickers.
Hardware →
Finde dein lokales Coding-Setup
Tell us the machine you have—or the band you will buy. We match stacks and models that fit. Share the URL; the recommendation travels with it.
Query string updates as you change options.
FAQ
Is self-hosting always cheaper than an API?
No. Local token stickers are ~$0, but you pay electricity, depreciation, and ops. Use the break-even jobs/month estimate on this page.
Why can Q4 weights fit 24GB but long context fail?
KV cache grows with context. Weights fitting the card is not the same as native 128K–262K fitting. Budget KV separately.
Should I start with vLLM?
Not on a laptop. Start with Ollama or llama.cpp (MLX on Apple). Use vLLM/SGLang on a dedicated NVIDIA box when you need server throughput.
Does AI Hippo global rank measure coding skill?
No. The unified board is a context/price-style proxy. Vendor SWE tables are not independent harness reruns unless stated.
Related insights
Assumptions and sources
Hardware bands are classes, not street quotes. VRAM estimates use curated GGUF footprints plus size-class KV factors. API job cost uses OpenRouter primary listing when present. choose-ai-infrastructure-guide covers API hosts—not desk GPUs.