Choose by VRAM
Model weights, KV cache, context length, and runtime overhead all consume memory. A model that fits at 4K may not fit at 32K, so the recommendation recalculates as context changes.
LOCAL LLM · HARDWARE ATLAS
Which model is all hype — and which one fits your GPU?
Pick your hardware and workload. We put this week’s model heat against the evidence on your machine.
GB
·
AI HASHRATE / WEEKLY EVIDENCE
Heat makes a claim. Hardware gives the verdict.
OpenRouter heat is attention, not proof. This court checks the claim against your exact hardware, quantization, and context.
Other verdicts same rig, different evidence
Everything that fits
Nearly fits and the smallest change that fixes it
LOCAL LLM TOOL
Local LLM GPU and VRAM calculator
AI Hashrate helps you choose an open-weight model from the memory you actually have. Select a GPU or enter usable VRAM to compare fit, context, estimated tokens per second, and a runnable command.
Estimate real-world speed
The estimate uses memory bandwidth and model size. Measured anchors are shown separately from estimates, so tok/s is a planning number rather than a benchmark promise.
Run the selected model
After you choose a model, copy an exact command for llama.cpp, Ollama, MLX, or vLLM when the runtime is supported. Model IDs and filenames are never invented.
How to use the calculator
- 1. Set memory — detect your GPU or enter the usable VRAM shown by your operating system.
- 2. Tune the workload — choose chat, coding, reasoning, or the biggest model that fits, then set quantization and context.
- 3. Copy the command — use the recommended model and runtime command as a starting point for local inference.
Common questions
What does “fits” mean?
Fits means the estimated model weights, KV cache, and runtime overhead stay within the usable memory you entered. Leave headroom for the operating system and other applications.
Are the tok/s numbers benchmarks?
They are estimates unless a card and model combination is marked measured. Actual speed varies with drivers, backend, batch size, thermals, and prompt length.
Go deeper: local LLM hardware guides
These short guides explain the constraints behind the calculator and link back to real hardware entries in the catalog.