On a 16 GB GPU, 1 catalog model run fully on the GPU at an 8,192-token context. The most capable is Devstral-Small-2-24B-Instruct-2512 at IQ4_XS (needs ~14.9 GiB). Pick a smaller model or a lower quant for more headroom.
Figures assume a 16 GB GPU paired with 32 GB of system RAM (a typical desktop), judged at an 8,192-token context. Speed depends on the specific card — this page ranks by fit, not speed.
TL;DR
16 GB only just reaches the dedicated code-model class, so expect a very short list. The catalog's code specialists cluster in the low-20B range and up, and at 16 GB that class fits only at compact 4-bit quants with little to spare. Plenty of general instruct models handle code too — the full 16 GB list is the wider menu. For coding at this tier, budget context before anything else.
Coding is long-context work — repository files, diffs, error logs, and multi-turn sessions all pile into the window, and the KV cache grows to hold them. A specialist that fits at this page's stated context with thin margins will feel it the moment you feed it a real project, so the practical choices are: keep the window modest, step the quant down, or run a smaller general model with headroom instead. Check your intended context in the calculator before settling.
Devstral-Small-2-24B-Instruct-2512 at IQ4_XS · 24B params · needs ~14.9 GiB
One command to run it (llama.cpp):
llama-server -m Devstral-Small-2-24B-Instruct-2512-IQ4_XS.gguf -c 8192 -ngl 999| Model | Sweet-spot quant | Fits in |
|---|---|---|
| Devstral-Small-2-24B-Instruct-2512 mistralai | IQ4_XS Runs fully on GPU | ~14.9 GiB |
Derived live from the fit engine + catalog at an 8,192-token context. "Fits in" is the modelled VRAM the sweet-spot quant needs (weights + KV cache + overhead). Speed depends on your specific card — check a GPU page or the calculator.
Because it only lists models the catalog tags as code specialists that also run fully on a 16 GB GPU — and the specialist class starts right around the edge of what 16 GB holds. Plenty of general instruct models are competent at code; the general 16 GB page shows the full menu of what the card runs.
Yes, with discipline. The specialist class fits at compact quants with thin margins, so keep the context window realistic and watch memory. For long agent-style sessions, a smaller general model with real headroom is often more dependable than a specialist that barely fits — test both on your own codebase.
The usual advice — specialists for heavy coding, generalists for mixed work — tilts toward general models at 16 GB, because the specialists here fit with so little slack. A mid-size general model at a healthy quant with context room can beat a squeezed specialist in day-to-day use; the general 16 GB page lists those options.
More than for chat — files, diffs, and logs add up fast, and agent workflows compound it. At 16 GB, context is the scarcest resource: the KV cache comes out of the same budget as the weights, so size the window you'll actually use in the fit calculator, not the default.
You can, but interactive coding is the worst case for it — generation slows sharply once the model leaves the GPU, and editor plugins and agent loops multiply that wait across many calls. This page lists only full-GPU fits; if you want a bigger code model at full speed, that's what the larger VRAM tiers are for.
selected to compare · pick at least 2