Qwen3-VL MoE VRAM: 30B-A3B vs 235B-A22B

Short answer: Qwen3-VL’s MoE branch has two very different memory classes. The complete official Q4 paths are 19.269GB for 30B-A3B and 142.906GB for 235B-A22B after adding the Q8_0 projector. A3B and A22B describe routed computation, not the amount of model data that must be stored or placed.

Qwen3-VL 30B-A3B vs 235B-A22B

Model	Complete official Q4 + Q8 projector	Complete official Q8 + Q8 projector	Official BF16 weights	Practical starting tier
30B-A3B	19.269GB	33.196GB	62.142GB	32GB Q4; 48GB Q8; 80GB BF16
235B-A22B	142.906GB	250.692GB	471.340GB	192GB Q4; 320GB Q8; 640GB BF16

Active parameters are not stored-weight memory

Both official configurations declare 128 experts and select eight experts per token. Sparse routing limits which experts participate in a token’s computation, but it does not delete the other expert weights from the checkpoint. The A3B and A22B labels therefore cannot be used as 3B- or 22B-sized VRAM estimates.

Active parameters can affect compute behavior, but they do not replace direct artifact evidence. Hardware fit still depends on total weights, projector, cache, runtime implementation, vision workload, device placement and concurrency.

Cache geometry also differs

Model	Layers	KV heads	FP16/BF16 cache per token	32K payload	256K payload
30B-A3B	48	4	96KiB	3GiB	24GiB
235B-A22B	94	4	188KiB	5.875GiB	47GiB

Native 256K context is an architecture limit, not a promise that the allocation fits beside the model. Batch size, concurrency, cache precision and multimodal tokens change observed memory.

Choose the correct deployment class

• Choose 30B-A3B when 24GB–80GB hardware is the realistic range and the exact workload validates well enough.
• Choose 235B-A22B only when a 143GB-class Q4 deployment and explicit multi-device engineering are acceptable.
• Do not choose from active parameters alone: benchmark representative image, document and video tasks.
• Do not sum board labels casually: validate model partitioning and peak memory on every device.

Detailed Qwen3-VL MoE guides

The Qwen3-VL 30B-A3B guide covers 24GB through 80GB paths, complete projector pairs and a 96KiB-per-token cache. The Qwen3-VL 235B-A22B guide sums every split file and covers 160GB through 640GB multi-device paths.

Compare the dense Qwen3-VL family, general MoE memory guide and cross-family multimodal guide.

Reproducible MoE vision test

1. Record repository, revision, variant, all language shards, projector and checksums.
2. Record runtime/version, MoE kernels and whether image/video inputs are actually active.
3. Record every device, capacity, topology, parallelism mode and offload target.
4. Record weight, projector and cache precision plus per-device placement.
5. Fix image dimensions/count, video sampling, text allocation, output target and concurrency.
6. Measure accelerator and system memory per device before load, after load, after vision encoding and at peak generation.
7. Report preprocessing time, time to first token, throughput, communication/fallback behavior and errors.

Use the VRAM calculator, multi-GPU guide, workload test and benchmark submission.

Sources and limits

• Qwen official 30B-A3B repository
• Qwen official 30B-A3B GGUF repository
• Qwen official 235B-A22B repository
• Qwen official 235B-A22B GGUF repository
• Qwen3-VL official code repository
• HighVRAM methodology

Reviewed October 25, 2026. Artifact sizes and architecture fields were read from official Qwen repositories. HighVRAM has not independently benchmarked every topology, runtime, MoE kernel, variant or multimodal workload. File size is not peak resident memory.

Qwen3.5 MoE path: the 35B-A3B VRAM guide explains why 3B active parameters do not reduce stored-weight memory to a 3B model.