Best Mini PC for Running Ollama Under $300 (2026)

Most mini PCs under $300 are sold on cores. Cores are the least important number in this price band. A small model generating text on a CPU is not limited by how many ALUs you bought; it is limited by how fast the machine can stream four and a half gigabytes of weights through memory, once per token. That is a bandwidth problem, and bandwidth is the first thing a cheap listing hides.

So this guide is not a spec-sheet ranking. It is the shortlist we would buy from, plus the checks that stop you paying for a machine that will disappoint you on day one.

What we measured, and what we did not

Be clear about the difference, because most "best mini PC" pages are not:

• Measured on our own hardware: memory throughput, wall power, and CPU temperature under sustained load. Those numbers are below, with dates.
• Arithmetic from those measurements: what that throughput implies for token rates. It is labelled as arithmetic because it is.
• Vendor claims, unverified by us: the CPU model, core count, RAM type, port layout and included accessories on the three specific listings below. We check the mechanics of a listing (price behaviour, variant traps, order counts) but we have not had these boxes on a bench.

We would rather write that paragraph than pretend otherwise. If you want the measured side of this site, start with the SSD guide, where every number is ours.

The three bands worth considering

Around the bottom of the range. This is the band where a box is honest about being small: a low-power SoC, 8 GB, a 256 GB drive, one Ethernet port. It will run a 3B-class model at a speed you can live with, serve files, and sip power. It will not run a 7B model comfortably, and no amount of tweaking changes that — see the RAM floor below.

Around the middle. This is where 16 GB and, sometimes, two memory slots appear. For most readers this is the correct band: 16 GB is the difference between "a 7B model fits" and "a 7B model fits and the OS still has room to breathe", and a second SODIMM slot is the difference between one memory channel and two.

The top of the budget. More cores, more storage, a better NIC, sometimes a marginally nicer chassis. You get real benefits here — a 2.5G port, an extra drive bay, quieter fans — but you do not automatically get twice the token rate, because the ceiling is still the memory bus. A more expensive box with a single SODIMM is slower at inference than a cheaper one with two.

The three listings we watch in these bands are in the cards at the bottom of this page. Prices there carry the date we checked them, because they move.

The number that decides everything: memory bandwidth

A 7B model quantised to 4-bit is roughly 4.4 GB on disk. To emit one token, the machine has to read essentially all of those weights. So the hard ceiling on token rate is:

memory bandwidth ÷ 4.4 GB, in tokens per second

We measured the numerator on our own box — a 2011 mobile dual-core, i5-2520M, one DDR3-1600 SODIMM, which is exactly the class of hardware this site is about:

A copy loop is a fair proxy: it is the same "read a lot, write almost nothing" pattern an inference kernel has, and on two old cores it lands at roughly 41% of theoretical peak. At 5.2 GB/s, streaming a 4.4 GB model takes about 0.85 seconds — so a couple of tokens per second, at best, before any compute overhead. That is arithmetic, not a benchmark, and a tuned kernel will beat a naive copy. But the order of magnitude is the point.

Now the part that actually helps you shop: that number is mostly a property of the memory bus, not the CPU. Single-channel DDR3-1600 peaks at 12.8 GB/s. Dual-channel DDR3-1600 doubles it. Dual-channel DDR4-3200 is 51.2 GB/s — four times our box. Same arithmetic, same model, four times the ceiling.

Which means, in this budget:

1. Two SODIMM slots populated beat a better CPU. Always.
2. Soldered single-channel RAM caps the machine forever. No upgrade path, no fix.
3. A "16 GB" listing that ships one 16 GB stick is not the same product as "16 GB" as 2×8 GB. Same capacity, half the bandwidth. Ask, or check the photos for two slots and two sticks.

If a listing will not tell you how many slots it has, it is telling you something.

The RAM floor is 8 GB, and 16 GB is the useful number

The 4.4 GB of weights is not the whole footprint. Add the KV cache (up to a gigabyte at a few thousand tokens of context), the model loader, and an operating system with its page cache. Round it up and:

• 4 GB: a 7B model does not fit. It will not "run slowly"; it will fail to load or thrash swap. We run a 4 GB machine here as a NAS and we do not pretend otherwise — see what a 4 GB NAS can actually do, where the answer is "be a NAS, run small models, and stop there".
• 8 GB: the honest floor for a 7B model. It fits, the machine is usable, and you will notice the OS competing with the model. Fine as a dedicated box that does nothing else while you chat.
• 16 GB: the comfortable number. Room for the model, room for the page cache so the second load is instant instead of a two-minute wait, room to run the rest of your stack.
• 32 GB: only worth it if you intend to run 13B-class models or a bigger context. Bandwidth, not capacity, becomes the limit again.

One storage detail people forget: pulling a model goes over the network. On a 100 Mbps link — which is what our older boxes actually negotiate — we measured 11.4 MB/s sustained. A 4.4 GB model is therefore about six and a half minutes of transfer before you even start. On gigabit it is under a minute. If you are choosing between a nicer CPU and a gigabit port on a machine that will pull models regularly, take the port.

What a cheap box actually does under sustained load

We stress-tested our own hardware rather than trusting a cooler's marketing. Every figure below is measured, on the same machine, from the same sensors:

Two things fall out of that table.

First, these boxes throttle, and they throttle immediately. One core was enough to hit the thermal limit in three seconds, and the chip then sat below its own base clock — meaning it lost roughly a fifth of its rated performance to cooling that was sized for idle. A "faster CPU" in the same chassis buys you nothing you can feel; it just hits the wall harder. When a listing shows a desktop-class chip with a token heatsink and no fan, that is not a bargain, it is a space heater with a network port.

Second, the whole machine costs less to run than a light bulb. 20 W idle, 40 W flat out. Two of these plus a switch is under 100 W. That is the real argument for small hardware: not that it is fast, but that it is cheap enough to leave on for years.

What to check on the listing before you pay

The mechanics of these listings matter more than the spec sheet. In order of how much money they cost you when you get them wrong:

1. Memory channels and slot count. Two slots, both populated, or nothing. Soldered is a permanent cap.
2. Barebones or complete. Some prices are for a chassis with no RAM and no storage. By the time you add both, the "cheap" box costs more than the complete one next to it. Read the variant name, not the headline price.
3. "RAM/ROM" in the title. That phrasing comes from Android TV boxes. It usually means eMMC storage, not an SSD, and eMMC is what makes a machine feel broken.
4. NIC speed. 100 Mbps versus 1 Gbps is a ten-fold difference in how long a model pull takes. It is also the port you will still be using in five years.
5. Drive bays. One M.2 slot is enough to boot; a 2.5" SATA bay is what lets you put cheap bulk storage in later. If the SATA is provided by a bridge chip rather than the chipset, expect it to run warm.
6. Cooling described in words, not adjectives. A fan and a fin stack, or a 6 W part with a passive heatsink. Anything else in this price band is a compromise and should be priced like one.
7. Who honours the warranty. These are usually grey imports. A 90-day DOA window from a seller who answers messages is worth more than a longer promise from a seller who does not.

What we would not buy

• A fanless box with a desktop-class CPU in it. The physics does not work and we measured why.
• 4 GB soldered, at any price.
• Anything marketing "AI TOPS" while omitting the RAM spec — the TOPS number describes an NPU that Ollama will not use, while the RAM is what your model actually has to live in.
• A "mini PC" whose only storage is eMMC.
• A box with no second memory slot when a same-priced one has two. That is a permanent 2x on the one number that matters.

What we are still measuring

This page will get better, and it will say when. In progress: token rates per band on the same prompt and the same model, 8 GB versus 16 GB at identical clock speeds, and single-channel versus dual-channel on the same box. Those are bench numbers with a stopwatch behind them, not arithmetic.

Until then the buying rule is short enough to remember: buy two memory channels, 16 GB if you can, a gigabit port, and a fan. The CPU badge on the front is the least informative thing in the listing.

If you are still deciding whether local models are worth the electricity at all, read Ollama first — it is a one-line install, and the honest answer is that a small model on the hardware you already own is a different tool from a frontier model in the cloud, not a slower copy of one.