Qwen3 · Apache-2.0 · HF: Qwen/Qwen3-30B-A3B
Qwen3 30B-A3B — VRAM, KV Cache & Hardware Matrix
Serverless API Rate
$0.12 in / $0.50 out per 1M
Parameters
30.5B (3.3B act.)
Attention
GQA (4 KV / 32 Q)
Layers × Hidden
48 × 2048
Max Native Context
40k tokens
MoE Experts
8 / 128
Quantization Memory Footprint
| Quantization | BPW | Weights Only | Total @ 8k Ctx | Total @ 32k Ctx | Quality |
|---|---|---|---|---|---|
| FP16 | 16.00 | 57.95 GiB | 59.73 GiB | 61.98 GiB | Lossless Reference |
| Q8_0 | 8.50 | 30.78 GiB | 32.57 GiB | 34.82 GiB | Near-Lossless |
| Q6_K | 6.57 | 23.79 GiB | 25.58 GiB | 27.83 GiB | High Precision |
| Q5_K_M | 5.69 | 20.61 GiB | 22.39 GiB | 24.64 GiB | Balanced High |
| Q4_K_M | 4.85 | 17.57 GiB | 19.35 GiB | 21.60 GiB | Recommended 4-bit |
| MLX_4BIT | 4.50 | 16.30 GiB | 18.08 GiB | 20.33 GiB | Apple Silicon Optimized |
| IQ4_XS | 4.25 | 15.39 GiB | 17.17 GiB | 19.42 GiB | Compact 4-bit |
| Q3_K_M | 3.91 | 14.16 GiB | 15.94 GiB | 18.19 GiB | Aggressive Compression |
Hardware Compatibility & Speed Leaderboard (Q4_K_M @ 32k Context)
| Hardware | Status | Total / Limit | Decode | TTFT (4k) | Max Safe Ctx | Inspect |
|---|---|---|---|---|---|---|
| Cloud NVIDIA H200 SXM (141 GB) | COMFORTABLE | 21.6 / 140.4 GiB | 674.3 tok/s | 0.08s | 40k | Simulate → |
| Cloud NVIDIA H100 SXM (80 GB) | COMFORTABLE | 21.6 / 79.4 GiB | 470.6 tok/s | 0.08s | 40k | Simulate → |
| Cloud NVIDIA A100 SXM (80 GB) | COMFORTABLE | 21.6 / 79.4 GiB | 279.6 tok/s | 0.19s | 40k | Simulate → |
| NVIDIA RTX PRO 6000 Blackwell (96 GB) | COMFORTABLE | 21.6 / 95.0 GiB | 251.7 tok/s | 0.23s | 40k | Simulate → |
| NVIDIA RTX 5090 (32 GB) | COMFORTABLE | 21.6 / 31.0 GiB | 245.8 tok/s | 0.28s | 40k | Simulate → |
| Cloud NVIDIA L40S (48 GB) | COMFORTABLE | 21.6 / 47.4 GiB | 118.5 tok/s | 0.18s | 40k | Simulate → |
| Apple Mac Studio M3 Ultra (256 GB) | COMFORTABLE | 21.4 / 199.7 GiB | 106.8 tok/s | 1.03s | 40k | Simulate → |
| Apple Mac Studio M3 Ultra (512 GB) | COMFORTABLE | 21.4 / 399.4 GiB | 106.8 tok/s | 1.03s | 40k | Simulate → |
| 2x NVIDIA RTX 3090 Rig (48 GB) | COMFORTABLE | 21.6 / 46.6 GiB | 105.2 tok/s | 0.97s | 40k | Simulate → |
| Apple Mac Studio M2 Ultra (192 GB) | COMFORTABLE | 21.4 / 149.8 GiB | 104.4 tok/s | 1.14s | 40k | Simulate → |
| Apple Mac Studio M4 Max (64 GB) | COMFORTABLE | 21.4 / 49.9 GiB | 70.3 tok/s | 1.56s | 40k | Simulate → |
| Apple Mac Studio M4 Max (128 GB) | COMFORTABLE | 21.4 / 99.8 GiB | 70.3 tok/s | 1.56s | 40k | Simulate → |
| Apple MacBook Pro M3 Max (64 GB) | COMFORTABLE | 21.4 / 49.9 GiB | 50.8 tok/s | 2.08s | 40k | Simulate → |
| Apple MacBook Pro M3 Max (128 GB) | COMFORTABLE | 21.4 / 99.8 GiB | 50.8 tok/s | 2.08s | 40k | Simulate → |
| Apple Mac Mini M4 Pro (48 GB) | COMFORTABLE | 21.4 / 37.4 GiB | 33.8 tok/s | 3.19s | 40k | Simulate → |
| Apple Mac Mini M4 Pro (64 GB) | COMFORTABLE | 21.4 / 49.9 GiB | 33.8 tok/s | 3.19s | 40k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (128 GB) | COMFORTABLE | 21.4 / 96.0 GiB | 31.7 tok/s | 1.17s | 40k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (64 GB) | COMFORTABLE | 21.4 / 48.0 GiB | 31.7 tok/s | 1.17s | 40k | Simulate → |
| Ryzen AI 9 HX 370 Mini-PC (64 GB) | COMFORTABLE | 21.4 / 48.0 GiB | 14.1 tok/s | 3.28s | 40k | Simulate → |
| NVIDIA RTX 4090 (24 GB) | TIGHT | 21.6 / 23.0 GiB | 138.2 tok/s | 0.35s | 40k | Simulate → |
| NVIDIA RTX 3090 (24 GB Used) | TIGHT | 21.6 / 23.0 GiB | 125.2 tok/s | 0.80s | 40k | Simulate → |
| AMD Radeon RX 7900 XTX (24 GB) | TIGHT | 21.6 / 23.0 GiB | 122.0 tok/s | 0.56s | 40k | Simulate → |
| NVIDIA RTX 5080 (16 GB) | OOM | 21.6 / 15.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 5070 Ti (16 GB) | OOM | 21.6 / 15.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 5060 Ti (16 GB) | OOM | 21.6 / 15.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 3060 (12 GB) | OOM | 21.6 / 11.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 (16 GB) | OOM | 21.4 / 10.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 (24 GB) | OOM | 21.4 / 18.0 GiB | OOM | — | 12k | Simulate → |
Interactive Qwen3 30B-A3B Simulator
30.5B (3.3B active) · GQA
1792 GB/s · 250 TFLOPS
3. Weight Quantization
4.85 bpw
4. KV Cache
32,768 tok
Compute Roofline
COMFORTABLE FIT Memory-Bandwidth Bound
Max Safe Ctx: 40k
Memory Allocation Footprint 21.60 / 95.00 GiB usable
Weights: 17.57 GiB
KV Cache: 3.00 GiB
Overhead: 1.03 GiB
Headroom: 73.40 GiB
Decode Speed
251.7
tok/s (batch=1)
Prefill Speed
20705
prompt tok/s
TTFT Latency
0.23s
for 4,096 tok
Active Params
3.3B
of 30.5B total
llama.cpp Server Command
llama-server -hf Qwen/Qwen3-30B-A3B:Q4_K_M -ngl 99 -fa on -c 32768 --port 8080| Context | KV Cache | Total VRAM | Decode | Status |
|---|---|---|---|---|
| 4k | 0.38 GiB | 18.83 GiB | 542.2 tok/s | COMFORTABLE |
| 8k | 0.75 GiB | 19.35 GiB | 465.4 tok/s | COMFORTABLE |
| 16k | 1.50 GiB | 20.10 GiB | 362.8 tok/s | COMFORTABLE |
| 32k | 3.00 GiB | 21.60 GiB | 251.7 tok/s | COMFORTABLE |
Buy-vs-Cloud-vs-API Monthly TCO Break-Even
USE SERVERLESS APIOwn Hardware/mo
$161.09
17.2h active + power
Cloud GPU/mo
$36.64
@$1.85/hr spot/pod
Serverless API/mo
$13.50
Payback: 941.7 mo
Recommended Hardware & Cloud Deployment
Contextual to fit statusEngineering FAQ — Qwen3 30B-A3B
How much VRAM does Qwen3 30B-A3B require in Q4_K_M?
Qwen3 30B-A3B weights occupy 17.57 GiB in Q4_K_M. Including 32k FP16 KV cache and runtime buffers, total memory is 21.60 GiB.
What is the cheapest local hardware that runs Qwen3 30B-A3B?
NVIDIA RTX PRO 6000 Blackwell (96 GB) (96 GB) runs Qwen3 30B-A3B Q4_K_M at ~251.7 tok/s decode and 0.23s TTFT.
How does Qwen3 30B-A3B scale KV cache at long context?
Qwen3 30B-A3B uses GQA with 4 KV heads across 48 layers (128 head_dim).