Arynwood/local-ai-benchmarks

Local AI benchmarks: how fast do local language models run on ordinary hardware?

Measured Ollama generation speeds on real machines, CPU only and on a GPU, plus tool-calling and decision-routing tests for local agents. Every row comes from a recorded run of the open bench.py scripts by Arynwood Technology.

Short answers from this data

• On a 2015 laptop (Intel i3-5005U, 6 GB RAM), TinyLlama ran at 12.9 tokens/s on Linux Mint 22.3 and 9.4 tokens/s on Windows 10, same Ollama 0.11.4 and model digest: 37% higher on Mint for this setup. Write-up, DOI 10.5281/zenodo.23290682.
• An RTX 3060 12 GB ran Llama 3.2 3B at 77.7 tokens/s against 12.8 on the 2012 Xeon CPU of the same machine (6.1x). Models up to 14B fit on the card. What can an RTX 3060 12GB run?
• Without a GPU, that 2012 Xeon ran Llama 3.2 3B at 12.8 tokens/s and Qwen2.5 7B at 6.4; most people read about 5 tokens/s. Local AI without a GPU

Speed at a glance

Warm generation speed is the median eval_count / eval_duration of three warm runs answering the same prompt with a 200-token cap, temperature 0 and a fixed seed.

Date	CPU	OS	Model	Quant	Mode	Warm tok/s	Ollama
2026-09-22	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	tinyllama	Q4_0	cpu	27.1	0.11.4
2026-09-22	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	llama3.2:3b	Q4_K_M	cpu	12.8	0.11.4
2026-09-22	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	qwen2.5:7b-instruct	Q4_K_M	cpu	6.4	0.11.4
2026-09-24	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	hermes3:8b	Q4_0	cpu	5.7	0.11.4
2026-09-24	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	qwen2.5-coder:14b	Q4_K_M	cpu	3.5	0.11.4
2026-09-24	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	gpt-oss:20b	MXFP4	cpu	2.4	0.11.4
2026-09-24	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	deepseek-coder:33b	Q4_0	cpu	1.5	0.11.4
2026-09-22	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	tinyllama	Q4_0	gpu	197.4	0.11.4
2026-09-24	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	llama3.2:3b	Q4_K_M	gpu	77.9	0.11.4
2026-09-22	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	llama3.2:3b	Q4_K_M	gpu	77.7	0.11.4
2026-09-24	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	mistral:latest	Q4_K_M	gpu	57.8	0.11.4
2026-09-24	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	llama3:8b-instruct-q4_0	Q4_0	gpu	50.3	0.11.4
2026-09-24	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	hermes3:8b	Q4_0	gpu	49.9	0.11.4
2026-09-24	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	qwen2.5:7b-instruct	Q4_K_M	gpu	49.0	0.11.4
2026-09-22	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	qwen2.5:7b-instruct	Q4_K_M	gpu	48.9	0.11.4
2026-09-24	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	qwen2.5-coder:14b	Q4_K_M	gpu	28.5	0.11.4
2026-09-24	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	gpt-oss:20b	MXFP4	gpu	6.5	0.11.4
2026-09-24	Intel Xeon E5-2665	Ubuntu 24.04.5 LTS	deepseek-coder:33b	Q4_0	gpu	2.7	0.11.4
2026-10-05	Intel Xeon Gold 6138 (virtual server)	Ubuntu 24.04.5 LTS	gemma3:1b	Q4_K_M	cpu	9.3	0.15.2
2026-10-05	Intel Xeon Gold 6138 (virtual server)	Ubuntu 24.04.5 LTS	llama3.2:1b	Q8_0	cpu	7.2	0.15.2
2026-10-05	Intel Xeon Gold 6138 (virtual server)	Ubuntu 24.04.5 LTS	gemma3:1b	Q4_K_M	cpu	5.8	0.35.1
2026-10-05	Intel Xeon Gold 6138 (virtual server)	Ubuntu 24.04.5 LTS	llama3.2:1b	Q8_0	cpu	5.0	0.35.1
2026-10-05	Intel Xeon Gold 6138 (virtual server)	Ubuntu 24.04.5 LTS	granite3.3:2b	Q4_K_M	cpu	3.7	0.15.2
2026-10-05	Intel Xeon Gold 6138 (virtual server)	Ubuntu 24.04.5 LTS	gemma:2b-instruct-q4_0	Q4_0	cpu	3.3	0.15.2
2026-10-05	Intel Xeon Gold 6138 (virtual server)	Ubuntu 24.04.5 LTS	gemma:2b-instruct-q4_0	Q4_0	cpu	3.1	0.35.1
2026-10-05	Intel Xeon Gold 6138 (virtual server)	Ubuntu 24.04.5 LTS	llama3.2:3b	Q4_K_M	cpu	2.8	0.15.2
2026-10-05	Intel Xeon Gold 6138 (virtual server)	Ubuntu 24.04.5 LTS	granite3.3:2b	Q4_K_M	cpu	2.8	0.35.1
2026-10-05	Intel Xeon Gold 6138 (virtual server)	Ubuntu 24.04.5 LTS	llama3.2:3b	Q4_K_M	cpu	2.5	0.35.1
2026-10-10	Intel i3-5005U	Linux Mint 22.3	tinyllama	Q4_0	cpu	12.9	0.11.4
2026-09-29	Intel i3-5005U	Windows 10	tinyllama	Q4_0	cpu	9.4	0.11.4

Tables in this dataset

• speed (40 rows): Generation speed: warm tokens per second and cold first-answer time, one row per model and mode
• tool_calling (1039 rows): Tool calling: one row per test case and run, with the tools called and the reply
• decision_routing (306 rows): Decision routing: one row per routing case and run for small decision models
• app_evals (504 rows): Arynwood MCP live evals: one row per eval test

raw/ holds every original file unchanged, with the per-machine reports. Pick a table in the viewer above or load one:

from datasets import load_dataset
speed = load_dataset("Arynwood/local-ai-benchmarks", "speed", split="train")

Add your machine

Run the script and open a results issue; accepted results land here automatically.

curl -O https://raw.githubusercontent.com/Arynwood-Technology/local-ai-benchmarks/main/bench.py
python3 bench.py llama3.2:3b --csv my-results.csv     # add --modes cpu without an NVIDIA GPU

Submit your results

Method and limits

• Prompt: "Explain in about 150 words why someone might run a language model locally."
• One cold request with the model unloaded, then warm requests. Cold time includes model load.
• CPU-only runs set num_gpu to 0. gpu_share_pct is how much of the loaded model sat on the GPU.
• One machine is one data point. Instruction sets, memory bandwidth, background work, power settings and runtime versions all change speed. Each source file's report in raw/ lists its conditions.

License and citation

Data: CC BY 4.0. Credit "Arynwood" and link to https://arynwood.com/local-ai-cpu-only.html. Scripts: MIT, in the GitHub repository. Citation details are in CITATION.cff. This mirror is rebuilt from GitHub on every new result.

Want local AI set up on your own machine? See Arynwood MCP and local AI setup services.