Tested 2026-06-16 - Docker 8GiB - CPU-only - Ollama 0.30.8

Best Local LLMs for 8GB RAM CPU-Only in 2026: Ollama Benchmarks

I tested six local LLMs inside an Ollama Docker container capped at 8GiB RAM with no GPU passthrough. The short answer: start with qwen2.5:0.5b or gemma3:1b. Use 3B models only if you can wait. 7B models can load, but they feel slow on this class of machine.

Last updated

2026-06-16

What changed

This article now includes raw Docker/Ollama evidence, measured tokens per second, memory pressure notes, FAQ, and a direct methodology link.

Review methodology

Fastest

qwen2.5:0.5b

58.02 tokens/s

Best range

0.5B to 3B

Smooth enough for local chat

Slow edge

qwen2.5:7b

7.15 tokens/s

How to read this test

Use the test order, not just the raw speed number.

The host CPU was fast, so an older 8GB laptop will usually produce lower absolute tokens/s. The useful signal is the order, memory pressure, and where the experience stops feeling comfortable. That is why the practical pick stays small: start below 3B, then treat 7B as a boundary test.

Open the 8GB model guide Compare your machine

Fit first

If the model leaves almost no memory headroom, normal apps can push the machine into swapping.

Comfort matters

A slower but smarter model is not always better if every short reply takes patience.

Retest locally

Use one prompt from your own workflow before making any model your daily default.

Field notes from the run

The main lesson: the best 8GB model is the one you keep using.

I do not read this benchmark as a contest to find the biggest model that loads. On an 8GB RAM CPU-only machine, the better question is which model gives useful answers without making the computer feel occupied by the model.

The fastest model is a setup check

qwen2.5:0.5b is the model I would use to prove Ollama, memory limits, and CPU-only inference are working before judging local AI quality.

1B felt like the practical daily range

The 1B-class models kept enough speed headroom that short prompts still feel casual instead of turning into a benchmark chore.

3B is where patience starts to matter

qwen2.5:3b was usable, but it is no longer instant. I would test my own prompts before calling it a default on an 8GB laptop.

7B is a boundary, not a recommendation

The useful lesson from qwen2.5:7b is that loading is not the same as comfort. It proves the edge, but it is not where I would start.

Quick answer

What should you run on an 8GB RAM CPU-only computer?

Install first

qwen2.5:0.5b is the safest first model. It used about 1GiB loaded memory and generated 58.02 tokens/s in this test.

Daily lightweight chat

gemma3:1b, llama3.2:1b, and qwen2.5:1.5b stayed above 25 tokens/s and are still comfortable.

Quality tradeoff

qwen2.5:3b is the upper practical range here. It ran at 16.77 tokens/s, which is usable but no longer snappy.

Raw Docker inspect and stats output proving the Ollama container used an 8GiB memory limit with no GPU request — Step 1: raw Docker inspect and stats output shows the 8GiB memory cap, no CPU cap, and GPURequests=null.

Raw Ollama log showing CPU inference compute and zero VRAM for the 8GB RAM CPU-only benchmark — Step 2: the Ollama log reports id=cpu, library=cpu, and total_vram="0 B".

Docker setup

Yes, I set up Docker first

Docker is not required if you only want to run Ollama on your own computer. For this article, Docker matters because it lets the test mimic an 8GB RAM CPU-only box while the host machine has more memory and a GPU. The container was launched with an 8GiB memory limit and no GPU request.

docker run -d --name aijupyter-8gb-cpu-ollama \
  --memory=8g --memory-swap=8g \
  -p 11435:11434 \
  -e OLLAMA_NUM_PARALLEL=1 -e OLLAMA_KEEP_ALIVE=0 \
  ollama/ollama:latest

The important proof line is GPURequests=null. Ollama also logged id=cpu library=cpu and total_vram="0 B", so the benchmark did not use the host RTX GPU.

Raw Ollama model list output showing six local models pulled inside the 8GiB CPU-only container — Step 3: docker exec ollama list confirms the six models tested inside the same container.

Raw benchmark JSON output with measured tokens per second for six local LLMs on 8GB RAM CPU-only Docker — Step 4: the benchmark JSON records eval_count, eval_duration, wall time, and docker memory stats.

Measured speed

Tokens per second: 0.5B and 1B models are the real 8GB sweet spot

I used the same prompt, num_ctx=2048, num_predict=128, and temperature=0 for every model. Tokens per second were calculated from Ollama's returned eval_count / eval_duration.

Model	Size	Tokens/s	Loaded memory	Verdict
qwen2.5:0.5b	397 MB	58.02	1003MiB / 8GiB	Best first install
gemma3:1b	815 MB	30.70	2.132GiB / 8GiB	Recommended
llama3.2:1b	1.3 GB	25.89	3.295GiB / 8GiB	Usable
qwen2.5:1.5b	986 MB	25.65	3.88GiB / 8GiB	Usable
qwen2.5:3b	1.9 GB	16.77	4.45GiB / 8GiB	Usable, not snappy
qwen2.5:7b	4.7 GB	7.15	6.409GiB / 8GiB	Slow boundary

Raw benchmark JSON is available at benchmark-results.json. The host CPU was an AMD Ryzen 7 9800X3D, so older 8GB laptops should expect lower absolute numbers, but the model order is still useful.

Raw qwen2.5 7B benchmark evidence showing 7.15 tokens per second and 6.409GiB memory use — Step 5: qwen2.5:7b loaded, but the raw result shows 7.15 tokens/s and 6.409GiB of 8GiB used.

What gets slow

7B works, but it is the wrong default for 8GB RAM CPU-only

The 7B result is the useful warning. It did not crash. It did not require a GPU. But it took 30.07 seconds to generate 128 tokens and used 6.409GiB inside the 8GiB container. On a real 8GB laptop with a browser, editor, and OS services open, that margin can disappear quickly.

Keep context short

Use 2048 tokens or less when memory is tight. Bigger context windows increase KV cache memory and make a barely usable model feel worse.

Avoid 13B+ on this tier

If 7B is already at 7.15 tokens/s on this CPU, larger dense models are not a good daily-driver target for an 8GB RAM CPU-only machine.

Install commands

The install order I would use

First

ollama run qwen2.5:0.5b

Fastest, lowest memory, best sanity check.

Second

ollama run gemma3:1b

Still fast enough for a small private assistant.

Stretch

ollama run qwen2.5:3b

Better capability tradeoff, but expect slower replies.

FAQ

Practical 8GB local LLM questions

What is the best local LLM for 8GB RAM with no GPU?

In this Docker test, qwen2.5:0.5b was the safest first install at 58.02 tokens/s, while gemma3:1b was still comfortable at 30.70 tokens/s. If you want a stronger model and can tolerate waiting, qwen2.5:3b was usable at 16.77 tokens/s.

Can an 8GB RAM CPU-only computer run a 7B local LLM?

Yes, qwen2.5:7b loaded in the 8GiB container, but it only generated 7.15 tokens/s and used 6.409GiB of the 8GiB memory limit. That is usable for experiments, but too slow for a smooth daily assistant.

Do I need Docker to test local models on an 8GB machine?

Docker is not required for normal Ollama use, but it is useful for this benchmark because it enforces the 8GiB RAM ceiling and proves the test was CPU-only. For daily use, installing Ollama directly is simpler.

Will these tokens per second match an old 8GB laptop?

No. The container was limited to 8GiB RAM, but the host CPU was an AMD Ryzen 7 9800X3D. Older laptop CPUs will usually be slower, so treat the order as a fit and relative-speed guide rather than a universal speed promise.

What local LLM size should I install first on 8GB RAM?

Install a 0.5B or 1B model first, then move to 3B only if the laptop stays responsive. Starting small proves the runtime works before you spend time downloading a model that may feel too slow.

Is 8GB RAM enough for coding with a local LLM?

It is enough for tiny private helpers, short explanations, and simple snippets. Repository-wide coding, long files, test loops, and agent workflows usually need a larger machine or a hosted coding model.

More local LLM hardware guides

Use the test order, not just the raw speed number.

Fit first

Comfort matters

Retest locally

The main lesson: the best 8GB model is the one you keep using.

What should you run on an 8GB RAM CPU-only computer?

Install first

Daily lightweight chat

Quality tradeoff

Yes, I set up Docker first

Tokens per second: 0.5B and 1B models are the real 8GB sweet spot

7B works, but it is the wrong default for 8GB RAM CPU-only

Keep context short

Avoid 13B+ on this tier

The install order I would use

Practical 8GB local LLM questions

What is the best local LLM for 8GB RAM with no GPU?

Can an 8GB RAM CPU-only computer run a 7B local LLM?

Do I need Docker to test local models on an 8GB machine?

Will these tokens per second match an old 8GB laptop?

What local LLM size should I install first on 8GB RAM?

Is 8GB RAM enough for coding with a local LLM?

Best Local LLMs for 8GB RAM in 2026: Ollama, LM Studio, and Model Picks

Best Local LLMs for 16GB RAM in 2026: Ollama and LM Studio Picks

Best Local LLMs for 32GB RAM in 2026: 7B, 8B, and 14B Picks

Best Local LLMs for RTX 5090 in 2026: 32 GB VRAM Picks

Best Local LLMs for RTX 5080 in 2026: 16 GB VRAM Picks

Best Local LLMs for RTX 5070 Ti in 2026: 16 GB VRAM Picks

Best Local LLMs for RTX 5070 in 2026: 12 GB VRAM Picks

Best Local LLMs for RTX 5060 Ti 16GB in 2026: 16 GB VRAM Picks

Best Local LLMs for RTX 5060 Ti 8GB in 2026: 8 GB VRAM Picks

Best Local LLMs for RTX 5060 in 2026: 8 GB VRAM Picks

Best Local LLMs for RTX 4090 in 2026: 24GB VRAM Picks

Best Local LLMs for RTX 4080 in 2026: 16 GB VRAM Picks

Best Local LLMs for RTX 4070 Ti Super in 2026: 16 GB VRAM Picks

Best Local LLMs for RTX 4070 Ti in 2026: 12 GB VRAM Picks

Best Local LLMs for RTX 4070 in 2026: 12 GB VRAM Picks

Best Local LLMs for RTX 4060 Ti 16GB in 2026: 16 GB VRAM Picks

Best Local LLMs for RTX 4060 Ti 8GB in 2026: 8 GB VRAM Picks

Best Local LLMs for RTX 4060 in 2026: 8 GB VRAM Picks

Best Local LLMs for RTX 3090 in 2026: 24 GB VRAM Picks

Best Local LLMs for RTX 3080 in 2026: 10 GB VRAM Picks

Best Local LLMs for RTX 3070 in 2026: 8 GB VRAM Picks

Best Local LLMs for RTX 3060 Ti in 2026: 8 GB VRAM Picks

Best Local LLMs for RTX 3060 in 2026: 12 GB VRAM Picks

Best Local LLMs for MacBook in 2026: Apple Silicon Picks