GPUFits
Models / Qwen3 30B-A3B
Qwen3 30B-A3B VRAM Requirements & GPU Pairing
Qwen3 30B-A3B is lesson one of MoE magic: 30.5B total parameters (128 experts, 8 active per token) but only 3.3B active — VRAM follows the total (Q4_K_M @8K ~21GB, tight on 24GB), while speed follows the 3.3B active params: a theoretical ~374 tok/s on an RTX 4090, over 4× a same-size dense model. That is the MoE law: capacity by total params, speed by active params.
Apache-2.0, native 32K context (YaRN to 131K), thinking-mode support. Against its dense sibling Qwen3 32B: much faster, slightly less VRAM, benchmark quality trading blows — dense partisans credit 32B with more solid knowledge, MoE partisans cannot refuse 4× speed. MoE's edge widens further under CPU offload (each token reads only the 3.3B-active slice from system RAM). For a responsive general assistant on a 24GB card, this is our top recommendation.
Architecture Specs
| Total parameters | 30.5B |
| Active parameters(MoE: VRAM follows total params, speed follows active params) | 3.3B |
| Layers | 48 |
| KV heads | 4 |
| Head dim | 128 |
| Native context | 32K |
| License | apache-2.0 |
| KV bytes per token (fp16) | 96.0 KB |
VRAM by Quant Tier (@8K context, fp16 KV)
| Quant | Weights | Total (incl. 1.5GB runtime overhead) |
|---|---|---|
| Q8_0 | 32.4 GB | 34.7 GB |
| Q6_K | 25.0 GB | 27.4 GB |
| Q4_K_M | 18.7 GB | 21.0 GB |
| Q3_K_M | 15.3 GB | 17.6 GB |
| Q2_K | 12.1 GB | 14.4 GB |
KV cache and runtime overhead do not change across quants; longer contexts grow the KV part linearly — adjust context length in the GPU checker tool.
GPU Verdict Matrix (Q4_K_M @8K)
| GPU | Usable VRAM | Verdict | Recommended quant | Theoretical speed | |
|---|---|---|---|---|---|
| RTX 3060 12GB | 12.0 GB | Not feasible | needs 2 cards | — | Try it → |
| RTX 3090 | 24.0 GB | Tight fit | Q4_K_M | ≈347 tok/s | Try it → |
| RTX 4070 Ti Super | 16.0 GB | Not feasible | Q2_K | ≈385 tok/s | Try it → |
| RTX 4090 | 24.0 GB | Tight fit | Q4_K_M | ≈374 tok/s | Try it → |
| RTX 5090 | 32.0 GB | Comfortable | Q6_K | ≈496 tok/s | Try it → |
| RTX A6000 | 48.0 GB | Comfortable | Q8_0 | ≈164 tok/s | Try it → |
| A100 80GB | 80.0 GB | Comfortable | FP16 | ≈232 tok/s | Try it → |
| H100 80GB | 80.0 GB | Comfortable | FP16 | ≈381 tok/s | Try it → |
| RX 7900 XTX | 24.0 GB | Tight fit | Q4_K_M | ≈356 tok/s | Try it → |
| Mac mini M4 Pro (48GB) | 36.0 GB | Comfortable | Q8_0 | ≈58 tok/s | Try it → |
| Mac Studio M4 Max (64GB) | 48.0 GB | Comfortable | Q8_0 | ≈117 tok/s | Try it → |
| Mac Studio M3 Ultra (96GB) | 72.0 GB | Comfortable | FP16 | ≈93 tok/s | Try it → |
Theoretical speed = bandwidth × 0.75 ÷ per-token weight bytes (active params for MoE); real-world results vary with framework/driver/CPU, ±30%.
FAQ
- Why is Qwen3 30B-A3B so fast?
- MoE: only 8 of 128 experts fire per token — 3.3B active parameters. Decode reads ~2GB of weights per token, so an RTX 4090 theoretically hits ~374 tok/s versus ~38 tok/s for dense 32B.
- Can a 24GB GPU run Qwen3 30B-A3B?
- Tight fit. Q4_K_M @8K is ~21GB with ~3GB to spare. Q6_K (~27.4GB) does not fit — that needs a 32GB card (RTX 5090).
- Qwen3 30B-A3B or Qwen3 32B?
- Pick 30B-A3B for speed and CPU-offload friendliness; pick 32B for dense-model knowledge consistency and predictable behavior. On the same 24GB card, the former is “fast and good enough”, the latter “slower but more solid”.
Related guides
- MoE Model Hardware Requirements: Total vs Active Parameters
- Why Memory Bandwidth Determines LLM Inference Speed
- CPU Offload: Is Partial GPU Offloading Worth It?
Try Qwen3 30B-A3B in the GPU compatibility checker →
Data verified 2026-09-01