What runs on a 16 GB GPU: local AI models that fit
A 16 GB NVIDIA card runs more than its memory suggests. For mixture-of-experts models the FLE engine keeps the experts in RAM and only the busiest ones on the GPU, so the RAM decides which model fits. With 128 GB of RAM the biggest model in the catalog that fits a 16 GB card is Qwen3.8-Flash-Next.
16 GB card with 32 GB of RAM
| Model | Does it fit? | Measured on this hardware | One command |
|---|---|---|---|
| Llama 3.2 1B1B, a first try on any laptop | fits | not measured on this size | flelink run llama-3.2-1b |
| Katy 27B + DFlash2dense 27B with a draft model | fits | 19.4 tok/s16 GB card, 51 of 66 layers on it; one run, Ryzen 9 9950X3D with 128 GB RAM; a slower CPU will be slower | flelink run katy-27b-dflash |
| Kate 27B + DSpineKaty without the refusal direction | fits | 37.3 tok/s16 GB, 55 of 66 layers on the card; simulated on a 32 GB card in a Ryzen 9 9950X3D with 128 GB RAM; a slower CPU will be slower | flelink run kate-27b-dspine |
| Qwen3.8-Flash-Next125B mixture of experts, 88.9 GB | too big | not measured on this size | flelink run fle-iq3-m |
| DeepSeek V4.1 Flash + DSpark763B mixture of experts, 232 GiB | too big | not measured on this size | flelink run deepseek-v41-ista-dspark |
| Rin (DeepSeek V4.1 Flash, split)763B, hot half on the GPU | too big | not measured on this size | flelink run rin-deepseek-v41-dspark |
16 GB card with 64 GB of RAM
| Model | Does it fit? | Measured on this hardware | One command |
|---|---|---|---|
| Llama 3.2 1B1B, a first try on any laptop | fits | not measured on this size | flelink run llama-3.2-1b |
| Katy 27B + DFlash2dense 27B with a draft model | fits | 19.4 tok/s16 GB card, 51 of 66 layers on it; one run, Ryzen 9 9950X3D with 128 GB RAM; a slower CPU will be slower | flelink run katy-27b-dflash |
| Kate 27B + DSpineKaty without the refusal direction | fits | 37.3 tok/s16 GB, 55 of 66 layers on the card; simulated on a 32 GB card in a Ryzen 9 9950X3D with 128 GB RAM; a slower CPU will be slower | flelink run kate-27b-dspine |
| Qwen3.8-Flash-Next125B mixture of experts, 88.9 GB | fits | about 15 tok/s16 GB card, 64 GB RAM (8 Nov 2026) | flelink run fle-iq3-m |
| DeepSeek V4.1 Flash + DSpark763B mixture of experts, 232 GiB | too big | not measured on this size | flelink run deepseek-v41-ista-dspark |
| Rin (DeepSeek V4.1 Flash, split)763B, hot half on the GPU | too big | not measured on this size | flelink run rin-deepseek-v41-dspark |
16 GB card with 128 GB of RAM
| Model | Does it fit? | Measured on this hardware | One command |
|---|---|---|---|
| Llama 3.2 1B1B, a first try on any laptop | fits | not measured on this size | flelink run llama-3.2-1b |
| Katy 27B + DFlash2dense 27B with a draft model | fits | 19.4 tok/s16 GB card, 51 of 66 layers on it; one run, Ryzen 9 9950X3D with 128 GB RAM; a slower CPU will be slower | flelink run katy-27b-dflash |
| Kate 27B + DSpineKaty without the refusal direction | fits | 37.3 tok/s16 GB, 55 of 66 layers on the card; simulated on a 32 GB card in a Ryzen 9 9950X3D with 128 GB RAM; a slower CPU will be slower | flelink run kate-27b-dspine |
| Qwen3.8-Flash-Next125B mixture of experts, 88.9 GB | fits | not measured on this size | flelink run fle-iq3-m |
| DeepSeek V4.1 Flash + DSpark763B mixture of experts, 232 GiB | too big | not measured on this size | flelink run deepseek-v41-ista-dspark |
| Rin (DeepSeek V4.1 Flash, split)763B, hot half on the GPU | too big | not measured on this size | flelink run rin-deepseek-v41-dspark |
The verdicts are the rule FLE Link applies before it downloads anything (the rules in full). A speed is shown only for the hardware it was measured on. A size nobody measured yet? flelink bench --task pagoda --share posts yours to the benchmarks board.
Other sizes: 12 GB · 24 GB · 32 GB. Any other GGUF file works with flelink run /path/to/model.gguf; the calculator on the main page takes any size, an Apple Silicon Mac or no GPU.
Install FLE Link
curl -fsSL https://flelink.run/install.sh | shWindows: irm https://flelink.run/install.ps1 | iex. Open source, MIT. No account, no telemetry. What FLE Link does →