Local AI on 8GB in 2026: Model Sizes That Actually Fit (Ollama)

Last updated: September 9, 2026 · 5-minute read

The most load-bearing number in local AI is a boring one: file size. On September 9, 2026, Ollama's catalog lists Qwen 3.5's 2B at 2.7GB and Gemma 4's E4B QAT build at 6.1GB — and both numbers redraw the line for what an 8GB laptop can carry.

A rule before any table: this refresh publishes sizes, not benchmarks — every gigabyte below was read from the Ollama library pages on September 9, 2026, and nothing was re-measured for this piece. The math that follows is arithmetic on published sizes, stated as arithmetic. The earlier how-to on running Llama 3.1 on a used laptop stays the hands-on companion; this is the catalog catch-up, and the 8GB advice needed it.

What moved since the Llama era

Two families redrew the small-model shelf this year. Qwen 3.5 landed on February 16, 2026, with its flagship open-weights release at 397B parameters — and the small sizes below are part of the same family line. Gemma 4 followed on April 2, 2026, announced by Google as "byte for byte, the most capable open models", with the E2B and E4B positioned explicitly as edge builds.

The registry stats say these are not niche drops. As listed on September 9, 2026, the Qwen 3.5 page shows about 19.9 million pulls and the Gemma 4 page about 24.5 million. Both families advertise vision, tools, and thinking capabilities; Gemma 4 also lists audio.

The September 2026 size table

Here is what the new families actually weigh, exactly as the tags list it:

And here are the older baselines that the 8GB conversation was built on, same source, same date:

Below all of them, smollm2's 135m, 360m, and 1.7b tags cover the featherweight end for hardware weaker than anything discussed here.

Both tables stop where the 8GB story stops. The families run much higher — Qwen 3.5 lists sizes up to 122b, Gemma 4 goes 12b, 26b (including a 26b-a4b MoE), and 31b — and none of those belong in this laptop's budget, which is exactly why they are not in the table.

Reading Gemma 4's edge tags

The E2B and E4B names are Google's own edge-model sizing, and the QAT suffix marks its quantization-aware-trained builds — memory-efficient versions of the same models, per Google's model documentation. This post does not invent quantization detail beyond what the tags show.

One pattern in the table is worth pausing on, because it is the kind of thing file sizes expose. The QAT build of the bigger edge model — gemma4:e4b-it-qat at 6.1GB — is smaller than the default build of the smaller one, gemma4:e2b at 7.2GB. The tag choice matters more than the family name suggests, which is an inference from the listed sizes rather than a measured claim.

The 8GB arithmetic, honestly

The whole budget reduces to one line: model file, plus context, plus the operating system and everything else. On 8GB machines the picture is fuzzier than on desktops, because integrated graphics often carve out shared memory — so treat every budget below as an approximation, not a promise.

Worked as subtraction: qwen3.5:4b at 3.4GB leaves roughly 4.6GB for context, the OS, and a browser you swore you closed. gemma4:e4b-it-qat at 6.1GB leaves about 1.9GB — close-everything territory, and a non-starter on machines where the GPU reserves a larger shared slice. The 7.2GB default E2B build is effectively out of the 8GB conversation entirely, which is why the QAT tags exist.

The older baselines slot in cleanly against the new math. deepseek-r1:7b at 4.7GB — the old "smallest serious model" — now sits mid-pack, above the entire Qwen 3.5 small trio. OpenAI's open-weight gpt-oss family starts at a 20b tag, which is past an 8GB budget no matter how the arithmetic is arranged.

The honest unit of local AI in 2026 is the gigabyte, not the benchmark. Sizes are published; vibes are not.

License cards before you ship

Sizes are only half of due diligence. The Gemma 3 and Gemma 4 pages list Google's Gemma Terms of Use, and deepseek-r1 is listed under MIT — both machine-readable from the model pages themselves.

Qwen 3.5 is the honest flag: its page did not yield a clean, machine-readable license during this pass, so this post makes no claim about it. If Qwen 3.5 is heading into anything commercial, read the license card on the model page first — that is the whole advice, and it is the kind of sentence that saves a product later.

My updated 8GB advice

For a daily driver, qwen3.5:2b at 2.7GB is the new comfortable middle — small enough to leave room, new enough to be current. qwen3.5:0.8b at 1.0GB is the always-on option for modest hardware, and qwen3.5:4b at 3.4GB is the step up when output quality matters more than headroom.

If the multimodal capability list matters, gemma4:e2b-it-qat at 4.3GB is the realistic Gemma 4 entry point on 8GB, and the E4B QAT build belongs on a 16GB machine. The old baselines have not stopped working — llama3.2:3b and qwen3:4b are still fine, sensible pulls — they just stopped being the interesting part of the catalog.

TL;DR

• Qwen 3.5 (0.8b at 1.0GB, 2b at 2.7GB, 4b at 3.4GB) and Gemma 4 edge builds (E2B QAT at 4.3GB, E4B QAT at 6.1GB) redraw the 8GB shortlist, per ollama.com listings on September 9, 2026.
• The E4B QAT build is smaller than the default E2B build — the tag matters more than the family name.
• The 8GB math is arithmetic, not a benchmark: file size plus context plus OS, with shared-GPU memory as the wildcard.
• deepseek-r1:7b at 4.7GB is now mid-pack; gpt-oss starts above the budget entirely.
• Sizes are the only numbers here on purpose — no benchmark claims. The Qwen 3.5 license is unverified in this pass; check the model card before commercial use.

Keep reading

• The original how-to: Llama 3.1 on a used laptop
• Why local AI is a privacy decision first
• Privacy-first tools that work in India
• Local AI & Privacy — the topic hub

---

Not affiliated with Ollama, Google, Alibaba, or Meta. Sources: the Ollama model library pages for each model listed above, fetched September 9, 2026.