Starter kits
Everything you need. Nothing you don't.
Complete local-LLM starter kits at three budgets: a board from our data plus the power supply, cooling and storage it actually needs. Every speed links to its source, and every price is one we could source, so totals are approximate.
Best local-LLM starter kit under $150.
Only one kit in our data fits this budget with sourced prices.
Our pick
M5Stack LLM Module Kit
≈ $79.90
Board + required parts, sourced prices at time of writing.
- 4.63 tok/s Qwen2.5 1.5B · StackFlow,
W4A16 (range 2.55–4.63 across 3 runs) · source ↗
Out of stock at M5Stack when we checked. It runs models converted for its Axera NPU (M5Stack says off-the-shelf models cannot be used directly), and it is built to pair with an M5 host controller, which is not priced here. source ↗
The LLM Module Kit has 32GB of eMMC and a microphone and speaker on board, so no storage card is needed. source ↗
All M5Stack Module LLM runs ›
Best local-LLM starter kit under $250.
Ranked by the fastest sourced run on Llama 3.1 8B, the largest model every kit in this tier has a measured result for.
Our pick
Radxa ROCK 5B+ 16GB kit
≈ $198.79
Board + required parts, sourced prices at time of writing.
- 3.72 tok/s Llama 3.1 8B · RKLLM runtime 1.2.1,
w8a8 (RKLLM conversion) · source ↗ - 16.69 tok/s Qwen2.5 1.5B · RKLLM v1.3.1,
w8a8 · source ↗
RK3588 speeds in our data run on the NPU through Rockchip's RKLLM, mostly from Rockchip's own reference-board benchmark, not a ROCK 5B+ specifically. Models must be converted to RKLLM format. source ↗
All Rockchip RK3588 runs ›
Also fits
Raspberry Pi 5 8GB kit
≈ $242.99
Board + required parts, sourced prices at time of writing.
- 2.05 tok/s Llama 3.1 8B · Ollama,
Ollama default (range 1.99–2.05 across 2 runs) · source ↗ - 5.95 tok/s Llama 3.2 3B · distributed-llama v0.12.3,
Q40 (range 0.14–5.95 across 6 runs) · source ↗ - 11.73 tok/s Qwen2.5 1.5B · Ollama,
Ollama default (range 0.27–11.73 across 4 runs) · source ↗
The easiest software path: Raspberry Pi OS plus Ollama. Everything runs on the CPU.
All Raspberry Pi 5 runs ›
Best local-LLM starter kit under $500.
Ranked by the fastest sourced run on Llama 3.2 3B, the largest model every kit in this tier has a measured result for.
Our pick
Jetson Orin Nano Super kit
≈ $449.99
Board + required parts, sourced prices at time of writing.
- 43.1 tok/s Llama 3.2 3B · MLC,
INT4 (range 4.3–43.1 across 3 runs) · source ↗ - 9.37 tok/s Qwen2.5 1.5B · Ollama,
not reported (range 4.21–9.37 across 4 runs) · source ↗
The power supply is in the box. The fastest numbers are NVIDIA's own MLC INT4 results; independent Ollama runs are much slower (see the range). source ↗
NVIDIA ships the dev kit with its 19V power supply, so you do not need to buy one. Storage is not included: NVIDIA asks for a 64GB+ UHS-1 microSD card or an NVMe SSD. source ↗
All Jetson Orin Nano Super runs ›
Also fits
Raspberry Pi 5 16GB kit
≈ $372.99
Board + required parts, sourced prices at time of writing.
- 5.95 tok/s Llama 3.2 3B · distributed-llama v0.12.3,
Q40 (range 0.14–5.95 across 6 runs) · source ↗ - 11.73 tok/s Qwen2.5 1.5B · Ollama,
Ollama default (range 0.27–11.73 across 4 runs) · source ↗
Speeds shown are for the Pi 5 in general. In Jeff Geerling's tests the 16GB and 8GB boards ran Llama 3.1 8B at nearly the same speed; the extra RAM buys room for bigger models, not speed. source ↗
All Raspberry Pi 5 runs ›
Also fits
Raspberry Pi 5 + AI HAT+ 2 kit
≈ $422.99
Board + required parts, sourced prices at time of writing.
- 2.65 tok/s Llama 3.2 3B · hailo-ollama,
Hailo-compiled (range 1.01–2.65 across 3 runs) · source ↗ - 6.82 tok/s Qwen2.5 1.5B · per M5Stack page,
W4A8 (range 4.37–6.82 across 5 runs) · source ↗
The HAT runs Hailo-compiled models through hailo-ollama. In our sourced runs it generates more slowly than the Pi 5's own CPU on the same models, so read the sources before you buy it for LLM speed.
The AI HAT+ 2 uses the Pi 5's single PCIe port, so you cannot run it alongside the M.2 HAT+ / NVMe SSD without a PCIe switch board. Boot from microSD instead. source ↗
All Raspberry Pi 5 + AI HAT+ 2 runs ›
How we build these kits
Board prices come from our board price list; accessory prices come from Seeed Studio, Samsung and Arace Tech product pages, checked 2026-10-17. We don't show Amazon prices, because they change hourly. A kit's total adds only prices we could source and lists anything left out. Speeds are the best single sourced run for that board family, and setups differ between sources, so open the source before you compare two boards closely. Some links are affiliate links (see below); they never change the ranking.Have one of these boards? Submit your own run ›