FreeLLMRAM for local AI New
When a model does not fit in your graphics card, the rest goes into system RAM. With 32 GB, even a 12 GB card runs one of the best free chat models at a good speed. Past 64 GB, chat gets better only with giant models that write slowly. Estimates from our simulator.
RAM speed matters: a 12 GB card plus RAM
Tokens per second, estimated. Blue: DDR5. Orange: DDR4. The note on each bar names the model.
What this shows · How to read it
What it shows. Estimated speed of the best free chat model that writes at reading speed on an RTX 3060 12 GB plus each desktop RAM kit (Gemma 4 26B A4B, a mixture-of-experts model that does not fit in 12 GB alone).
How to read it. Longer bar, faster. DDR5 bars are longer than DDR4 ones because the part of the model in RAM is read faster. Four-module kits run at lower MT/s, so they are a little slower than two-module kits.
Every kit: what it lets you run
| RAM kit | Type | Speed | Bandwidth | Best free model at reading speed, with RTX 3060 12 GB | Best free model at reading speed, with RTX 4090 | Buy |
|---|---|---|---|---|---|---|
| No extra RAM the card alone | — | — | — | Nothing ranked fits | Gemma 4 31B 1443 Arena · 36 tok/s · fully on the GPU | |
| 32 GB 2 × 16 GB · Desktop |
DDR4 | 3200 MT/s | 51 GB/s | Gemma 4 26B A4B 1434 Arena · 35 tok/s · GPU + RAM Bigger but slow: Gemma 4 31B, 1443 Arena, 3.6 tok/s | Gemma 4 31B 1443 Arena · 36 tok/s · fully on the GPU |
Amazon ↗ |
| 64 GB 2 × 32 GB · Desktop |
DDR4 | 3200 MT/s | 51 GB/s | Gemma 4 26B A4B 1434 Arena · 35 tok/s · GPU + RAM Bigger but slow: Gemma 4 31B, 1443 Arena, 3.6 tok/s | Gemma 4 31B 1443 Arena · 36 tok/s · fully on the GPU |
Amazon ↗ |
| 128 GB 4 × 32 GB · Desktop 4 modules: check that your board supports 128 GB |
DDR4 | 3200 MT/s | 51 GB/s | Gemma 4 26B A4B 1434 Arena · 35 tok/s · GPU + RAM Bigger but slow: Gemma 4 31B, 1443 Arena, 3.6 tok/s | Gemma 4 31B 1443 Arena · 36 tok/s · fully on the GPU |
Amazon ↗ |
| 32 GB 2 × 16 GB · Desktop |
DDR5 | 6000 MT/s | 96 GB/s | Gemma 4 26B A4B 1434 Arena · 49 tok/s · GPU + RAM Bigger but slow: Gemma 4 31B, 1443 Arena, 5.9 tok/s | Gemma 4 31B 1443 Arena · 36 tok/s · fully on the GPU |
Amazon ↗ |
| 64 GB 2 × 32 GB · Desktop |
DDR5 | 6000 MT/s | 96 GB/s | Gemma 4 26B A4B 1434 Arena · 49 tok/s · GPU + RAM Bigger but slow: Gemma 4 31B, 1443 Arena, 5.9 tok/s | Gemma 4 31B 1443 Arena · 36 tok/s · fully on the GPU |
Amazon ↗ |
| 96 GB 2 × 48 GB · Desktop |
DDR5 | 6000 MT/s | 96 GB/s | Gemma 4 26B A4B 1434 Arena · 49 tok/s · GPU + RAM Bigger but slow: Gemma 4 31B, 1443 Arena, 5.9 tok/s | Gemma 4 31B 1443 Arena · 36 tok/s · fully on the GPU |
Amazon ↗ |
| 128 GB 2 × 64 GB · Desktop |
DDR5 | 5600 MT/s | 90 GB/s | Gemma 4 26B A4B 1434 Arena · 48 tok/s · GPU + RAM Bigger but slow: Gemma 4 31B, 1443 Arena, 5.6 tok/s | Gemma 4 31B 1443 Arena · 36 tok/s · fully on the GPU |
Amazon ↗ |
| 192 GB 4 × 48 GB · Desktop 4 modules: most home boards run them at 5200 or less |
DDR5 | 5200 MT/s | 83 GB/s | Gemma 4 26B A4B 1434 Arena · 46 tok/s · GPU + RAM Bigger but slow: Gemma 4 31B, 1443 Arena, 5.3 tok/s | Gemma 4 31B 1443 Arena · 36 tok/s · fully on the GPU |
Amazon ↗ |
| 256 GB 4 × 64 GB · Desktop 4 modules: needs a recent board that supports 256 GB |
DDR5 | 4800 MT/s | 77 GB/s | Gemma 4 26B A4B 1434 Arena · 44 tok/s · GPU + RAM Bigger but slow: GLM-5.3-Flash, 1471 Arena, 4.1 tok/s | Gemma 4 31B 1443 Arena · 36 tok/s · fully on the GPU Bigger but slow: GLM-5.3-Flash, 1471 Arena, 4.4 tok/s |
Amazon ↗ |
| 32 GB 2 × 16 GB SO-DIMM · Laptop / mini PC |
DDR5 | 5600 MT/s | 90 GB/s | Gemma 4 26B A4B 1434 Arena · 48 tok/s · GPU + RAM Bigger but slow: Gemma 4 31B, 1443 Arena, 5.6 tok/s | Gemma 4 31B 1443 Arena · 36 tok/s · fully on the GPU |
Amazon ↗ |
| 64 GB 2 × 32 GB SO-DIMM · Laptop / mini PC |
DDR5 | 5600 MT/s | 90 GB/s | Gemma 4 26B A4B 1434 Arena · 48 tok/s · GPU + RAM Bigger but slow: Gemma 4 31B, 1443 Arena, 5.6 tok/s | Gemma 4 31B 1443 Arena · 36 tok/s · fully on the GPU |
Amazon ↗ |
| 96 GB 2 × 48 GB SO-DIMM · Laptop / mini PC |
DDR5 | 5600 MT/s | 90 GB/s | Gemma 4 26B A4B 1434 Arena · 48 tok/s · GPU + RAM Bigger but slow: Gemma 4 31B, 1443 Arena, 5.6 tok/s | Gemma 4 31B 1443 Arena · 36 tok/s · fully on the GPU |
Amazon ↗ |
Best free model at reading speed: the highest-voted free chat model on Arena that fits in the card plus 80% of the RAM, at 4-bit with 8K context, and writes at least 10 tokens per second. “Bigger but slow”: a higher-voted model that also fits but writes slower. Speed: estimated tokens per second from our simulator, calibrated on 44 real llama.cpp runs. Bandwidth: MT/s × 8 bytes × 2 channels.
Affiliate links: as an Amazon Associate, FreeLLMArena earns from qualifying purchases made through the “Amazon ↗” links. The links do not change the order, the numbers or the estimates on this page, and the price is the same for you.
How system RAM helps
Size decides what fits
A model at 4-bit needs about 0.6 GB per billion parameters, plus the context. If your card has 24 GB and the model needs 60 GB, the other 36 GB must fit in RAM. We count only 80% of the RAM as usable, because the system and your apps need the rest.
Speed decides if it is usable
Every word the model writes reads all its active weights. The part in RAM is read at RAM speed: dual-channel DDR5-6000 gives about 96 GB/s, against about 1,000 GB/s on an RTX 4090. So the more of the model sits in RAM, the slower it writes.
Mixture-of-experts models love RAM
MoE models (for example Qwen3-30B-A3B) use only a few billion parameters for each word. They can sit mostly in RAM and still write fast, which is why a big RAM kit plus a mid-range card can run surprisingly large models.
Two modules, not four
Home PCs have two memory channels. Two big modules usually run at full speed; four modules often run slower, and some boards limit the total. Check your motherboard's supported memory list before buying.
Want your exact setup? Use the Run at home simulator: pick your card, your RAM and a model. Choosing a card first? See GPUs for local AI.
Share this page:XLinkedInRedditWhatsApp
Spotted a wrong number or a broken link? Report a mistake