Qwen3 / deployment worksheet

Dense models, one MoE trap, twelve memory targets.

Q: How much GPU memory does Qwen3 8B need?

Qwen3 8B has 8.2 billion parameters. Its raw INT4 weight floor is about 3.82 GiB. With one 32K BF16 KV cache and 20 percent runtime headroom, this page's planning estimate is 9.98 GiB. Qwen's published Transformers benchmark reports 11,001 MB for AWQ-INT4 at a 30,720-token input.

Q: How much GPU memory does Qwen3 14B need?

Qwen3 14B has 14.8 billion parameters. Its raw INT4 weight floor is about 6.89 GiB. With one 32K BF16 KV cache and 20 percent runtime headroom, this page's planning estimate is 14.27 GiB. Qwen reports 15,323 MB for AWQ-INT4 at a 30,720-token input.

Q: How much GPU memory does Qwen3 32B need?

Qwen3 32B has 32.8 billion parameters. Its raw INT4 weight floor is about 15.27 GiB. With one 32K BF16 KV cache and 20 percent runtime headroom, this page's planning estimate is 27.93 GiB. Qwen reports 27,718 MB for AWQ-INT4 at a 30,720-token input.

Q: Does Qwen3 30B-A3B only need memory for 3.3B active parameters?

No. The 30B-A3B model activates about 3.3 billion parameters per token, but the full 30.5 billion parameter checkpoint still has to be loaded. Its raw INT4 weight floor is about 14.20 GiB, and this page's one-sequence 32K planning estimate is 20.64 GiB.

Q: How does 128K context change Qwen3 VRAM requirements?

KV-cache memory grows linearly with cached context. Moving from 32K to 128K multiplies the cache by four: from 4.5 to 18 GiB for Qwen3 8B, 5 to 20 GiB for 14B, 8 to 32 GiB for 32B, and 3 to 12 GiB for 30B-A3B for one BF16 cache.

Compare Qwen3 8B, 14B, 32B, and 30B-A3B across INT4, INT8, and BF16—then check the estimates against Qwen's published runtime measurements.

Open the editable 32B model →

Quick answer / INT4 at 32K

The 32B dense model needs more memory than the 30B-A3B MoE—but not by much.

Active parameters influence compute. Total parameters determine the weight memory that must be loaded. These capacity estimates use official Qwen parameter and architecture counts.

Qwen3 8B · 1 × 16 GiB 9.98 GiB 3.82 weights + 4.50 KV + headroom → Qwen3 14B · 1 × 24 GiB 14.27 GiB 6.89 weights + 5.00 KV + headroom → Qwen3 32B · 1 × 48 GiB 27.93 GiB 15.27 weights + 8.00 KV + headroom → Qwen3 30B-A3B · 1 × 24 GiB 20.64 GiB 14.20 weights + 3.00 KV + headroom →

32K serving matrix

Precision changes the weights. Model shape fixes the cache.

Planning target = (raw weight floor + one full 32K BF16 KV cache) × 1.20 runtime headroom. GPU examples reserve only 90% of listed VRAM.

Swipe the matrix horizontally →

Model	Architecture	INT4 target	INT8 target	BF16 target
Qwen3 8B	36L · 8 KVH · 128D	9.98 GiB 1 × 16 GiB	14.56 GiB 1 × 24 GiB	23.73 GiB 1 × 48 GiB
Qwen3 14B	40L · 8 KVH · 128D	14.27 GiB 1 × 24 GiB	22.54 GiB 1 × 48 GiB	39.08 GiB 1 × 48 GiB
Qwen3 32B	64L · 8 KVH · 128D	27.93 GiB 1 × 48 GiB	46.26 GiB 1 × 80 GiB	82.91 GiB 2 × 48 GiB
Qwen3 30B-A3B	48L · 4 KVH · 128D	20.64 GiB 1 × 24 GiB	37.69 GiB 1 × 48 GiB	71.77 GiB 1 × 80 GiB

A real quantized artifact adds scales, metadata, and sometimes higher-precision modules. Treat the mathematical weight floor as an input—not a promise that an arbitrary checkpoint will fit.

Read the estimate

MoE saves compute, not weights.
The 30B-A3B checkpoint loads all 30.5B parameters while activating about 3.3B per token.

Concurrency multiplies cache.
Every fully allocated sequence repeats the KV-cache budget.

Capacity is not speed.
A model that fits can still miss latency and throughput targets.

First-party runtime check

Qwen measured the same breakpoints.

Qwen3 8B · AWQ-INT4
11,001 MB at 30,720 input tokens

Qwen3 14B · AWQ-INT4
15,323 MB at 30,720 input tokens

Qwen3 32B · AWQ-INT4
27,718 MB at 30,720 input tokens

The official Transformers benchmark generated another 2,048 tokens and used one H20 96 GB GPU where possible. Runtime, kernels, quantization format, and allocation policy explain why observed memory is not identical to a planning formula.

The 128K context jump

Four times the context means four times the KV cache.

Qwen3 8B
4.5 GiB at 32K → 18 GiB at 128K

Qwen3 14B
5 GiB at 32K → 20 GiB at 128K

Qwen3 32B
8 GiB at 32K → 32 GiB at 128K

Qwen3 30B-A3B
3 GiB at 32K → 12 GiB at 128K

The official model cards describe 32,768 native context and 131,072 tokens with YaRN. Confirm the serving engine's actual cache allocation before reserving hardware.

Primary-source inputs

Copy the architecture. Check it against the runtime.

Parameter counts and model shapes come from Qwen's official model cards and configurations. Observed memory comes from Qwen's own Transformers benchmark—not a third-party VRAM table.

Open the GPU memory calculator → Open the KV-cache calculator → Compare the previous Qwen2.5 family → Official Qwen3 8B model card → Official Qwen3 14B model card → Official Qwen3 32B model card → Official Qwen3 30B-A3B model card → Official Qwen3 speed and memory benchmark →

How much GPU memory does Qwen3 8B need?

The 8.2B checkpoint has a 3.82 GiB INT4 weight floor. This page's one-sequence 32K planning target is 9.98 GiB; Qwen's AWQ-INT4 benchmark measured 11,001 MB at a 30,720-token input.

How much GPU memory does Qwen3 14B need?

The 14.8B checkpoint has a 6.89 GiB INT4 weight floor. Its 32K BF16 KV cache is 5 GiB, producing a 14.27 GiB planning target with 20% runtime headroom.

How much GPU memory does Qwen3 32B need?

The 32.8B checkpoint has a 15.27 GiB INT4 weight floor. Its 32K BF16 KV cache is 8 GiB, producing a 27.93 GiB planning target with 20% runtime headroom.

Does Qwen3 30B-A3B only need memory for 3.3B active parameters?

No. Active parameters describe per-token compute, while all 30.5B parameters must be loaded. The model's one-sequence 32K INT4 planning target is 20.64 GiB.

How does 128K context change Qwen3 VRAM requirements?

KV-cache memory grows linearly with context. Moving from 32K to 128K multiplies each model's cache by four before concurrency and runtime headroom.