How Much RAM Do You Need for Local AI? Testing 7 Models on a 2011 Processor
Authorship and approval
Published and updated
Key Takeaways
- We ran seven language models without a graphics card on a 2011 Intel i5-2400. They used between 2.9 and 8.5 GiB of memory.
- A larger model did not prove to be a better choice. Qwen 3.5 9B used twice as much memory as the 4B version while running slower, and it made the same mistakes.
- Gemma 4 E2B was the fastest, but it was the only one to write code with a syntax error.
- Without the word "AI" in the prompt, four models mistook a vector database for an algebra concept. After adding context, all seven answered correctly.
- Four out of seven got the expansion of the name NASK wrong. We would not publish texts from local models without checking them first.
What We Tested On
This is the server where our laboratory runs:
- Intel Core i5-2400 processor: 4 cores, 3.1 GHz, no AVX2,
- 15 GB of DDR3 memory,
- no graphics card for calculations,
- Ubuntu 26.04 and Ollama 0.40.7.
Every model received the same tasks: explaining a concept without context and with context, an email to a client, a short Python function, and a paragraph based on news about OpenAI's cooperation with NASK. Temperature 0, 4096 context tokens, one attempt per task, thinking mode disabled. During the test, the rest of the laboratory was also running on the server.
Results
Model | RAM | Loading | Short answers | Paragraph from text | Code |
|---|---|---|---|---|---|
Gemma 4 E2B | 6.3 GiB | 9 s | 8.7 tok/s | 57 s | syntax error |
Qwen 3.5 4B | 2.9 GiB | 11 s | 5.0 tok/s | 118 s | works, "cat," counts separately |
Nemotron 3.5 Bolt E4B | 6.6 GiB | 16 s | 4.7 tok/s | 110 s | works correctly |
Qwen 3.5 9B | 5.8 GiB | 18 s | 3.0 tok/s | 181 s | works, "cat," counts separately |
Ornith 1.5 9B | 5.8 GiB | 16 s | 3.0 tok/s | 207 s | works correctly |
Bielik 11B v2.3 | 7.2 GiB | 18 s | 2.2 tok/s | 314 s | works, "cat," counts separately |
Ministral 3 14B | 8.5 GiB | 20 s | 1.5 tok/s | 382 s | works, "cat," counts separately |
The models wrote paragraphs of varying lengths, so the time does not compare speed alone. We ran the code on the sentence "ala ma kota, kota ma ala". Swap memory did not grow in any test.
You can download the measurement data as a CSV file (CC BY 4.0).
We also tried Granite Switch 4.1 8B. It downloaded, but our Ollama did not accept it through regular chat, so it is not in the table.
What Surprised Us
A single word changed the answer. When asked what a vector database is, Qwen 4B, Qwen 9B, Ornith, and Bielik described an algebra space database. This is a correct answer to a different question. Nemotron and Ministral immediately answered about data for AI, and Gemma generally. After adding "in AI systems," all seven wrote about the right thing. A local model needs context in the prompt. We described the AI meaning in the AI Glossary.
The institution's name proved harder than summarization. The source provided the English name NASK. The models created Polish expansions: "National Institute of Scientific Research," "Technical Research Institute," "National Research Institute," and Ministral wrote "National Research Institute in the Tribune." The content of the summaries itself was closer to the source than those names.
The Polish model was not fast, but it did not make up the name. Bielik wrote cautiously "a research institute from Poland." Its paragraph took over 5 minutes to generate, so on this hardware, it is more like background work. In the final conclusion, however, it added something of its own: companies' access to NASK's expertise, which the source did not mention.
Is It Worth Buying a Graphics Card?
Not for us for now. An older test showed Gemma 12B Q4 running at about 1.2 tok/s and pointed to DDR3 memory as the bottleneck. We considered a GTX 1070 back then, but its 8 GB of memory was not enough for a 12B model with full context, so we postponed the purchase.
An old processor is enough when the work can wait. Now, Ornith tags topics for 2000 texts on it, taking about 45 seconds per text, which will take over a day. For live conversation with a large model, that is too slow.
Which Model to Use
- Low memory: Qwen 3.5 4B.
- Speed matters: Gemma 4 E2B, but check the code.
- Explanations and text: we would start with this after this test: Nemotron 3.5 Bolt E4B.
- In Polish, when time is not an issue: Bielik 11B. It is worth testing on your own texts.
With 8 GB of RAM, we would start with Qwen 4B. Models above 6 GiB might be too tight there, and Ministral definitely will not fit, but we did not test such a configuration.
What This Test Does Not Tell You
These are single attempts per task, on a single computer, and in Polish only. It shows specific answers rather than a general model ranking. We did not test other context sizes; a larger context takes up more memory.
Frequently Asked Questions
How much RAM is needed for local AI?
In our test, from 2.9 to 8.5 GiB for the model alone, with a context of 4096 tokens. On top of that, there is system memory and other programs.
Does local AI work without a graphics card?
Yes. On a 2011 processor, we achieved from 1.5 to about 9 tokens per second in short answers. This is enough for background tasks, but conversation is noticeably slow.
Which model was the fastest?
Gemma 4 E2B: 8.7 tokens per second and 57 seconds per paragraph. However, in the same test, it wrote code with a syntax error.
Do larger models give better answers?
In this test, not always. Qwen 9B and Ministral 14B were slower than Qwen 4B, and they also got the institution's name wrong.
Can a local model write blog posts on its own?
After this test, we would not do that. Four out of the seven models got the institution's name wrong. The text must be checked before publication.
Sources
Provenance manifest of this post (JSON) — which signals it was built from, what decisions the loops made, and when a human approved it.
Related posts
„AI Today” for visual reading. How we rebuilt Writer in one day
We added lab context and a scannable list to our daily AI Today review, winning 13:2 in reader tests.
AI Today: NVIDIA 6G and AMD InferaSim
Discover NVIDIA's 6G open tools growth, AMD's InferaSim LLM serving predictor, and the European Commission's TikTok DSA findings.
Small model versus Gemini: first trial of an open AI topic classifier
We built a first version of our AI text topic classifier, tested it against manual labels, and compared small models to Gemini.
AI Today: Google Genie 4 and Gemini 4 Argon
Discover the latest AI releases including Google Genie 4, Gemini 4 Argon, AMD ROCm 10.0.0, Liquid AI LFM3, and Anima's healthcare automation.