Local LLM
which model for your GPU or Mac?
Ollama, LM Studio, llama.cpp: pick your GPU (RTX, AMD, Apple Silicon) or enter VRAM. We list GGUF models that fit well, barely, or not at all — with max context and estimated tokens/sec.
Wondering which local LLM will run on your PC? VRAM and GPU bandwidth decide what actually fits (Ollama, LM Studio, GGUF). This tool estimates fit and speed before you download gigabytes of weights. See benchmarks · Guides
1. Hardware
System RAM enables offload (layers / KV cache). Multi-GPU: VRAM is summed, bandwidth includes overhead.
2. Filters
How to read “Fits well”, “Barely” and “Does not fit”?
Fits well: the model fits in VRAM with at least 4K context (and your filters). Barely: it fits but with little context, under your threshold, or via RAM offload. Does not fit: weights alone exceed available memory.
Where do estimated tok/s come from?
Token generation is mostly limited by GPU memory bandwidth. The tool estimates tok/s ≈ bandwidth / (weights + ~1 GB engine overhead). Without a listed GPU (manual VRAM), speed is not shown.
Quantization, KV cache, offload — short version
Quantization (Q4, Q8…) shrinks weight size. The KV cache grows with context. Offload sends some layers or cache to system RAM: the model runs, but often much slower (Ollama, LM Studio, llama.cpp).
Indicative estimates — real performance depends on runtime, drivers and load. GPU and GGUF catalogs are updated regularly. · Benchmarks · Models