On a 48 GB GPU, 2 catalog models run fully on the GPU at an 8,192-token context. The most capable is Qwen3-Coder-30B-A3B-Instruct at Q8_0 (needs ~34.6 GiB). Pick a smaller model or a lower quant for more headroom.
Figures assume a 48 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
48 GB is more memory than the current code-specialist class needs — which is its own kind of luxury. The catalog's code models run here at near-lossless quants with a large share of the card left over, so the budget goes to what coding workloads really eat: context, concurrency, or a second model alongside. It's a short list by design; the full 48 GB list shows everything the card runs, including much larger general models.
Qwen3-Coder-30B-A3B-Instruct at Q8_0 · 30.5B params · needs ~34.6 GiB
One command to run it (llama.cpp):
llama-server -m Qwen3-Coder-30B-A3B-Instruct-Q8_0.gguf -c 8192 -ngl 999| Model | Sweet-spot quant | Fits in |
|---|---|---|
| Qwen3-Coder-30B-A3B-Instruct Qwen | Q8_0 Runs fully on GPU | ~34.6 GiB |
| Devstral-Small-2-24B-Instruct-2512 mistralai | Q8_0 Runs fully on GPU | ~27.5 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 48 GB GPU — and the specialist class currently tops out well below what 48 GB holds. More memory doesn't add entries here; it adds headroom. The general 48 GB page shows the full menu of what the card runs.
Near-lossless quants as the default rather than a stretch, very long contexts for agent workflows and large files, and enough slack to serve parallel sessions or keep a second model loaded. For coding, that headroom tends to matter more day to day than a bigger model name would.
It's a real option — the general 48 GB tier fits much larger models than anything on this list, and strong general models are competent at code. Specialists still tend to lead on completion, editing, and tool-driven workflows. Test a big generalist against a specialist on your own codebase before deciding.
At this tier, usually yes. Memory is the only real budget: size each model in the fit calculator at its own context and make sure the combined footprint leaves margin. Two runtimes, or one server loading both models, both work — the KV cache of each session counts against the same card.
For fit, roughly — llama.cpp splits a model's layers across cards, so combined memory is the budget as a rule of thumb, minus a little per-card overhead. Interactive coding is latency-sensitive, though, and crossing between cards costs some speed, so expect a split rig to feel slower than one card with the same total memory.
selected to compare · pick at least 2