Llama 3.3 on NVIDIA GeForce RTX 4090 · 24GB VRAM

⚠️ Partial GPU offload — Llama 3.3 70B @ Q2_K
Llama 3.3 70B at Q2_K needs ~28.3 GB (weights 24.6 GB + KV 2.5 GB + overhead 1.2 GB @ 8K ctx) of your 21.6 GB usable VRAM — some layers spill to system RAM. Expect ~5–8 tok/s (slow).
context @ Q2_K: runs to 16K · won't fit past that on this rig
quantneedsspeed
✗ Q8_075.8 GB—
✗ Q5_K_M51.3 GB—
✗ Q4_K_M44.1 GB—
✗ Q3_K_M36.0 GB—
~ Q2_K28.3 GB~5–8 tok/s (slow)
$runlocal install llama3.3:70b
⚠ NVIDIA support is best-effort in v0.1 — verify before relying on it.
runlocal verdict card
download card share on x share on reddit drop it in your model-drop post
Llama 3.3: Partial GPU offload
for your own README — links back here

Same model, other machines: Apple M1 · Apple M1 Pro · Apple M2 · Apple M3 Pro · Apple M2 Pro · Apple M4 Pro

Also runs on a NVIDIA GeForce RTX 4090: DeepSeek-R1 (distill) · Qwen3 · Qwen2.5 · Gemma 3 · Gemma 4

Best models for a NVIDIA GeForce RTX 4090 · 24GB VRAM →

Architecturally similar

These share Llama 3.3's KV-cache geometry — they size memory the same way, so here's how they fit on a NVIDIA GeForce RTX 4090 · 24GB VRAM too:

modelsharesverdict here
Llama 3.1 8kv/128hd ✅ Runs comfortably
Qwen2.5-Coder 8kv/128hd ✅ Runs (tight)
Llama 3.2 8kv/128hd ✅ Runs comfortably
Mistral 7B 8kv/128hd ✅ Runs comfortably
Mistral Small 3 8kv/128hd ✅ Runs (tight)
Mistral Nemo 8kv/128hd ✅ Runs comfortably
Codestral 8kv/128hd ✅ Runs comfortably
Llama 3.2 Vision 8kv/128hd ✅ Runs comfortably

…and 8 more — see the Llama 3.3 page.

Check any combo yourself: open the checker →