Runyard tools
built around local LLM decisions.

Find exactly how much VRAM any model needs at any quantization level and context length.

See how larger context windows change KV cache growth, memory pressure, and practical fit.

Choose the right quant level for your hardware, quality target, and inference speed goals.

Match your GPU directly to the best-fit models and skip dead-end downloads.

Estimate what your chosen Ollama context setting means for memory use and local stability.

Estimate likely local inference speed before you commit to a model and runtime stack.

A guided compatibility check that funnels users to Runyard home for the live answer.

Convert model parameter counts and quant formats into memory estimates you can reason about.

Estimate whether partial GPU offload will help or just create a slow, awkward compromise.

Pick the right GGUF file variant without memorizing every quant suffix and packaging nuance.

A recommendation-first landing page that points users back to Runyard home for the live shortlist.

Count prompt size, estimate context consumption, and catch oversized inputs before they fail.

Compare model families across fit, speed, context, and practical use-case tradeoffs.

Estimate how much local inference can save compared with repeated paid API usage.

Estimate watts, monthly electricity cost, and BTU/hr heat for any local LLM rig.

Compare monthly OpenAI / Anthropic / Google API spend against a local GPU rig — find the break-even in months.

Estimate the on-disk size of any GGUF file by parameter count and quant format, plus VRAM at load.

Diagnose likely out-of-memory causes and suggest the fastest fixes for local inference failures.

QLoRA, LoRA and full fine-tuning memory for any model size, against your card.

Memory, fidelity and hardware for all three, at every model size.

How much memory is left for activations after the model, and what that buys.

Wall-clock estimate from dataset size, epochs and your hardware.

A complete, copy-pasteable command sized to your card and context.

Work out the fraction to set, and how many concurrent sequences it buys.

Concurrency from KV cache headroom, and what it does to per-user speed.

Cards required at each quantisation, and what that costs.

Disk for a library of models, before you run out mid-pull.

What a bigger card actually unlocks, in models and in speed.

Electricity plus hardware amortisation, expressed the way APIs price.

Break-even in months, from your real token volume.

Monthly and yearly electricity, from power draw and hours.

The hours per month at which renting stops being the cheaper option.

Monthly spend from requests, prompt length and price — with caching.

A one-off training cost against the per-request cost of retrieval.

Every catalogued card ranked by cost per usable gigabyte.

The lowest-cost card that fits, at every quantisation level.

What each memory configuration actually runs, and how fast.

The real cost of running a model that does not quite fit.

What CPU-only inference realistically gives you, at each model size.

Capacity, speed and cost across both, for the model you want to run.

Price per usable gigabyte against every alternative at the same capability.

The part of your VRAM that grows with every token in the conversation.

What --flash-attn buys you in context length, on your card.

How many chunks fit in your context once the prompt takes its share.

Disk and memory for a vector index, from document count and dimension.

Whether the index fits in memory, and what happens when it does not.

Lines of code to tokens, against real context windows.

Capacity follows total parameters; speed follows the active ones.

File size at each quantisation, and the wait on your connection.

Expected speedup from a draft model, at your acceptance rate.

Two cards compared on what actually matters: capacity and bandwidth.

Memory and real-time factor for each Whisper size, on your card.

Cards required to serve N people, from concurrency and daily volume.

Why a coding agent bills far more than the tokens you think you sent.