Start here

tensorless runs a model that was tuned for your GPU, on the engine it was tuned with, in one command. The first card is the NVIDIA Tesla P100 (16 GB), a server card that sells used for about $100. On it, tensorless runs a 27-billion-parameter model at 30 tokens a second.

curl -fsSL https://tensorless.dev/install.sh | sh
tensorless run iamMrDuncan/Qwen3.8-27B-PN3-GGUF

The launcher checks your card, picks the recipe for that model on that card, and pulls the engine image and the model file. It checks both by hash, then starts an OpenAI-compatible server on 127.0.0.1:8080. Every number on these pages was measured, and each one says how.

What you need

• An x86-64 Linux machine (the engine image is linux/amd64) with an NVIDIA Tesla P100 16 GB, free for this model alone.
• An x86-64 CPU with AVX2, FMA and F16C.
• An NVIDIA driver from R525 to R580. R580 is the last branch that supports the P100.
• podman or docker, with NVIDIA container support.
• About 13 GB of disk space for the model, plus a few GB for the engine image.

tensorless doctor checks the card, the driver, the container runtime, NVIDIA container support and disk space, and names the fix for anything missing. It does not check the CPU: grep -o -w -E 'avx2|fma|f16c' /proc/cpuinfo | sort -u should print all three.

Find your way

The docs follow Diátaxis: each page has one job.

TutorialsLearn by doingRun your first model on a P100, then measure your card against the receipt.How-to guidesGet one task doneInstall, prepare a host, run in the background, connect a client, benchmark, fix a failed start.ReferenceLook it upEvery command and flag, the recipe file, the engine image and its switches, the PN format, the receipts.ExplanationUnderstand whyHow the pieces fit, why the P100, where the engine came from, and why you can trust its numbers.

What it runs today

Launch command	Decode on one P100
tensorless run iamMrDuncan/Qwen3.8-27B-PN3-GGUF	30.03 tok/s
tensorless run iamMrDuncan/Qwen3.8-27B-UD-IQ3_XXS-Q3Floor-GGUF	26.18 tok/s

Both were measured on the published engine image with tensorless bench: the median of 3 runs of 256 greedy tokens after a 429-token prompt. The launch targets page has the details, and the receipts have every measurement.

Get help

Ask on Discord, or open an issue on GitHub. Results from your own P100 are welcome in both places.