Best GPU for Qwen3: VRAM Tiers for Every Model Size
As of 2026-10-07. Qwen is the most-run model family on the DiggTop board: 143 sourced community runs. Every number below comes from those runs — see benchmarks for sources and methodology.
TL;DR — pick by model size
| Qwen model class | VRAM needed | Our pick |
|---|---|---|
| Qwen 0.5B–4B (edge/speed) | 8–12GB | RTX 3060 12GB (~$495, 2026-09-28 snapshot) |
| Qwen 9B–14B | 12–16GB | RTX 3060 12GB / RTX 5060 Ti 16GB |
| Qwen 27B (3.8-27B, 3.6-27B) | 24GB | Used RTX 3090 (~$1,300, 2026-09-28 snapshot); RTX 5090 if you want NVFP4 speed |
| Qwen 35B-A3B (MoE, ~3B active) | 24–32GB | Radeon AI Pro R9700 32GB or any 24GB card |
| Qwen 70B-class | 48GB | 2× used 3090 or used A6000 48GB |
| Qwen3.8-Flash-Next 180B | 96GB+ | RTX PRO 6000 ×2, or Strix Halo 128GB at low quant |
Why Qwen is the community default
Qwen dominates our board: 143 of 336 sourced rows are Qwen-family. It is the family people actually run at home — Apache-licensed weights, a size for every budget, and first-class GGUF support in llama.cpp and LM Studio.
The VRAM picture
| VRAM tier | Qwen runs on board | Best tok/s |
|---|---|---|
| <24GB | 18 | 545.3 (Qwen3.5-0.8B @ MQ4, RX 9070 XT) |
| 24–48GB | 66 | 969.4 (Qwen3.6-35B-A3B @ MQ4R, Radeon AI Pro R9700) |
| >48GB | 59 | 1,060 (Qwen2.5-0.5B-Instruct @ Q8_0, RTX PRO 6000) |
The 24–48GB tier is where Qwen lives: 66 sourced runs. Its ceiling is set by a MoE rather than a dense 27B — 969.4 tok/s on a Radeon AI Pro R9700. (For reference, the fastest number anywhere on the board is 4,546.8 tok/s — a partial row: the source never stated an engine version, so it counts in nothing. The fastest sourced row is a small-model result, MobileLLM-125M at 1,271.4 tok/s on an RTX PRO 6000 Blackwell.)
27B: the community flagship
Qwen3.8-27B has 44 sourced runs — the most of any model. It fits 24GB at Q4, which makes the used RTX 3090 the default answer:
- Budget: used RTX 3090 24GB (~$1,300, 2026-09-28 snapshot) — Q4_K_M/IQ4_XS, interactive speeds
- Speed: RTX 5090 with NVFP4 — the fastest sourced 27B row is 435.1 tok/s (IQ4_XS,
lucebox, 131K context). A 4,546.8 tok/s NVFP4/sglang row exists in the database but is still partial, so it is not a board result - Multi-GPU budget route: Tesla V100 ×4 pulls 391 tok/s on the same model
35B-A3B MoE: the efficient middle
Qwen3.6-35B-A3B keeps only ~3B parameters active, so it runs comfortably on 24–32GB — and it flies on the Radeon AI Pro R9700 32GB (969.4 tok/s, MQ4R). If you want 35B-class quality without the 27B dense speed penalty, this is the one.
Going big: 70B and 180B
- 70B-class Qwen needs 48GB — the 48GB build guide covers dual-3090 vs A6000.
- Qwen3.8-Flash-Next (180B) needs 96GB+: RTX PRO 6000 ×2 on our board, or a Ryzen AI Max+ 395 with 128GB unified memory at low quant — see the Strix Halo guide.
Quant matters as much as the card
On Blackwell cards, NVFP4 is the fastest way to run 27B; on everything else, Q4_K_M and IQ4_XS are the workhorses. The full picture: quantization guide.
Before you buy
- Current $/GB ranking: value index
- Head-to-head tok/s: RTX 3090 vs RTX 5090 · model × hardware matrix
- Model × VRAM-tier matrix: compatibility matrix
- Every Qwen model page links its runs: start from benchmarks
Sources
- DiggTop community runs (143 Qwen-family, sourced): benchmarks · matrix
- Prices (dated snapshots): prices · value index
- Related: used GPU guide · 48GB build · quantization
Run counts and tok/s are sourced-only snapshots of the board on 2026-10-07; prices are the 2026-09-28 snapshot.