OnPrem World地端世界
中文

Qwen3-30B-A3B (2507)

Qwen3-30B-A3B (2507) by Alibaba Qwen (30.5B parameters, MoE with 3.3B active per token). At Q4_K_M with an 8K context it needs about 19.3 GiB of memory. A MoE with only 3.3B of its 30.5B parameters active per token, so it generates quickly even on mid-range hardware; it fits a 20 GB card at 4-bit.

Last updated: 2026-10-13

Model facts

DeveloperAlibaba Qwen
Parameters30.5B (3.3B active, MoE)[src]
Native context262,144 tokens[src]
LicenceApache-2.0[src]
Commercial useYes
KV cacheAll 48 layers (standard attention)[src]
Released2025-07

Good for: fast chat on modest GPUs · reasoning · multilingual · Chinese · tool use

Run it

ollama run qwen3:30b-instruct

Memory needed

Weights + KV cache + runtime overhead, in GiB. Lower quantization = less memory, slightly lower quality.

Quantization4K context8K context32K context
Q4_K_M19.0 GiB19.3 GiB21.6 GiB
Q5_K_M22.1 GiB22.5 GiB24.7 GiB
Q6_K25.4 GiB25.8 GiB28.1 GiB
Q8_032.6 GiB32.9 GiB35.2 GiB
FP1660.5 GiB60.9 GiB63.2 GiB

Which hardware runs it

Best quantization that fits each device with a 8K context. Estimates, not benchmarks. See the methodology.

HardwareMemoryBest quanttok/s (est.)
GeForce RTX 509032 GBQ6_K461
GeForce RTX 409024 GBQ5_K_M301
Radeon RX 7900 XTX24 GBQ5_K_M286
Radeon RX 7900 XT20 GBQ4_K_M280
GeForce RTX 309024 GBQ5_K_M279
RTX 6000 Ada Generation48 GBQ8_0192
RTX PRO 6000 Blackwell Workstation Edition96 GBFP16190
Radeon AI PRO R970032 GBQ6_K165
Arc Pro B7032 GBQ6_K157
RTX A600048 GBQ8_0153
RTX PRO 5000 72GB Blackwell72 GBFP16143
Arc Pro B6024 GBQ5_K_M136
M5 Ultra512 GBFP16118
M5 Max (18-core CPU / 32-core GPU)36 GBQ6_K110
M4 Max (14-core CPU / 32-core GPU)36 GBQ6_K98
M3 Ultra512 GBFP1681
M5 Max (18-core CPU / 40-core GPU)128 GBFP1660
M5 Pro64 GBQ8_057
M4 Max (16-core CPU / 40-core GPU)128 GBFP1654
M4 Pro64 GBQ8_051
M6 (24GB/32GB configurations, 170 GB/s)32 GBQ5_K_M47
M532 GBQ5_K_M42
RTX Spark (N1X) laptops & compact desktops128 GBFP1627
DGX Spark128 GBFP1625
Ryzen AI Max+ PRO 495 (Gorgon Halo) systems192 GBFP1625
Ryzen AI Max+ 395 (Strix Halo) systems128 GBFP1623
GeForce RTX 3060 (12 GB)12 GBDoesn't fit–
GeForce RTX 4060 Ti 16GB16 GBDoesn't fit–
GeForce RTX 4080 SUPER16 GBDoesn't fit–
GeForce RTX 5060 Ti 16GB16 GBDoesn't fit–
GeForce RTX 507012 GBDoesn't fit–
GeForce RTX 5070 Ti16 GBDoesn't fit–
GeForce RTX 508016 GBDoesn't fit–
Radeon RX 9060 XT 16GB16 GBDoesn't fit–
Radeon RX 9070 XT16 GBDoesn't fit–
Arc B58012 GBDoesn't fit–
M6 (16GB configuration, 153 GB/s)16 GBDoesn't fit–

Frequently asked questions

How much VRAM does Qwen3-30B-A3B (2507) need?

With an 8K context, about 19.3 GiB at Q4_K_M and 32.9 GiB at Q8_0 (weights + KV cache + runtime overhead). Longer contexts need more memory for the KV cache.

Can Qwen3-30B-A3B (2507) run on an GeForce RTX 4090?

Yes. Qwen3-30B-A3B (2507) fits on the GeForce RTX 4090 (24 GB) at Q5_K_M with an 8K context, using about 22.5 GiB, at an estimated 301 tokens/s.

Can I run Qwen3-30B-A3B (2507) on a Mac?

Yes. Qwen3-30B-A3B (2507) fits on the M3 Ultra (96 GB) at FP16 with an 8K context, using about 60.9 GiB, at an estimated 81 tokens/s.

Can Qwen3-30B-A3B (2507) be used commercially?

Qwen3-30B-A3B (2507) is released under Apache-2.0. The licence allows commercial use (read its conditions).

Sources