Rig Fit, now for expert streaming: size RAM and NVMe, not just VRAM

Every individual piece already exists. Calculators like llm-cal, LLMSizingTool and ModelFit cover sizing from theory, LocalScore and Llamabench cover crowdsourced tok/s, and vLLM bench serve and llama-benchy cover load testing. That makes a standalone load-test CLI or a standalone calculator a weak product. What I didn't find is anything that combines them. LocalScore's dataset is limited to one user on one GPU and has nothing on concurrency or serving engines. The calculators that do handle concurrency rely on bandwidth math, not measured results. None of them starts from team size or headcount. A realistic opening is a dataset of measured multi-user results, keyed on GPU, model, quant, engine, context length and concurrency. You would collect it with a thin wrapper around vLLM bench serve or a similar tool and use it to drive a recommender that answers questions like "8 people, 32k context, a Qwen-class 30B model". The output would be a parts list plus a launch config that has been tested on that hardware. The paid side most likely to work is installers and consultants: client-ready build reports, validated configs and a certified results badge. Technical founders will mostly use the free open data. The main risks are that LocalScore adds concurrency testing, or that vLLM or llm-cal ships a public results database. A crowdsourced dataset also has a cold-start problem, so you would probably need to fill it yourself by running cloud GPUs at the start.

• llm-cal

  The closest match on the sizing side. You give it a Hugging Face model and a GPU, and it estimates KV cache, fleet size, prefill and decode throughput, and concurrency limits, then writes a vLLM or SGLang launch command. All of this is calculated from theory. Nothing is measured, and it has no team-size input.

• LocalScore

  An open-source benchmark CLI with a public database of tok/s and TTFT results by hardware. It is the most established crowdsourced dataset I found, but it only tests one user on one GPU. It does not cover concurrency, multi-GPU setups or serving engines like vLLM.

• Llamabench

  Described as a crowdsourced platform that maps hardware, quantization and KV cache settings to tok/s. The only source I found was a startup directory listing, so I couldn't confirm how mature it is or how people submit results.

• vLLM bench serve

  vLLM's built-in load-test CLI, with --max-concurrency and request-rate controls, aimed at OpenAI-compatible servers. It covers the core load-testing feature for free but does not upload results anywhere.

• llama-benchy

  A CLI that benchmarks any OpenAI-compatible endpoint at different context depths. Its feature list mentions concurrency support. It also doesn't contribute results to a shared dataset.

• TrueFoundry LLM Benchmarking

  A hosted load tester that ramps up concurrent users and reports throughput, TTFT and inter-token latency. It runs inside their platform, so it isn't a local-first CLI.

• LLMSizingTool

  An open-source web tool that estimates VRAM, GPU count, latency and hardware cost (CAPEX) for a target queries per second (QPS). It is close to the build-recommender idea but works from estimates, not measurements.

• ModelFit calculator

  Checks whether a model fits in a given amount of VRAM and estimates speed. The site itself says its tok/s figures are estimated from memory bandwidth, not benchmarked.

• Inventive HQ LLM Inference Speed Calculator

  Estimates single-user tok/s from memory bandwidth, quantization and context length on consumer GPUs and Apple Silicon. It doesn't model concurrency.

• whichllm

  Ranks Hugging Face models against a user's own CPU, GPU and RAM, with estimated tok/s. It answers which model fits your hardware, the opposite question to which hardware fits your team.

• OpenClawDC local LLM hardware guide

  An editorial guide to hardware for 5, 10 and 25 users. It shows that installers and consultants are interested in this question, but it is static content, not a tool.

Checked 2026-10-11