Short answer: Qwen3-VL 235B-A22B is not a 22B memory model. The complete official Q4 path is 142.906GB with the Q8_0 projector. Treat 160GB as bounded, 192GB as practical for Q4, 256GB as too tight for Q8, 320GB as practical for Q8, and 512GB as tight for the official 471.340GB BF16 checkpoint.
A22B describes routed computation, not stored weights. All expert weights still require storage or placement. On multi-GPU systems, capacity is not automatically one pool: the runtime must partition the model, projector, cache and intermediate state across devices.
Official Qwen3-VL 235B-A22B file totals
| Official deployment path | Displayed files | Complete file size | Planning tier |
|---|---|---|---|
| Q4_K_M + Q8 projector | 3 language shards + 0.752GB projector | 142.906GB | 160GB bounded; 192GB practical |
| Q4_K_M + F16 projector | 3 language shards + 1.159GB projector | 143.313GB | 160GB bounded; 192GB practical |
| Q8_0 + Q8 projector | 6 language shards + 0.752GB projector | 250.692GB | 256GB too tight; 320GB practical |
| Q8_0 + F16 projector | 6 language shards + 1.159GB projector | 251.099GB | 256GB too tight; 320GB practical |
| Official BF16 safetensors | Checkpoint index totaling 471.340GB | 471.340GB | 512GB tight; 640GB safer |
Why A22B does not mean 22B storage
The official configuration declares 128 experts and routes eight experts per token. Sparse routing changes active computation; it does not discard the unselected expert weights from the checkpoint. The official Q4 language shards total 142.154GB before the projector, and the BF16 checkpoint totals 471.340GB.
Active parameters can describe part of the compute path, but artifact size, projector, cache, runtime buffers, image/video state and device placement determine memory planning. Fit does not establish speed, visual accuracy, output quality or cost effectiveness.
KV cache: 188KiB per allocated text token
Qwen’s official configuration lists 94 language layers, four KV heads and a 128-element head dimension. With FP16/BF16 keys and values, one sequence uses 2 × 94 × 4 × 128 × 2 = 192,512 bytes, or 188KiB per allocated text token.
| Allocated text tokens | FP16/BF16 KV payload |
|---|---|
| 8K | 1.46875GiB |
| 32K | 5.875GiB |
| 64K | 11.75GiB |
| 128K | 23.5GiB |
| 256K architectural maximum | 47GiB |
Multi-GPU capacity is not automatic pooling
- Partitioning: tensor, pipeline or expert parallelism must place weights and work across devices.
- Per-device headroom: one full machine total does not prove that every shard, layer or temporary allocation fits its assigned device.
- Interconnect: PCIe, NVLink/NVSwitch and host topology can change communication behavior; capacity alone does not predict speed.
- Projector placement: record where the vision encoder and its intermediate state reside.
- CPU offload: system RAM can extend fit but changes traffic and may change latency or throughput.
- Software support: verify the exact runtime, MoE kernels, quantization, model revision and device backend.
Vision and video add workload-dependent memory
The official GGUF deployment requires a 0.752GB Q8_0 or 1.159GB F16 projector. Image dimensions/count, video frames, duration, timestamp processing, DeepStack features, batching and concurrency add further allocations. A text-only response does not prove the multimodal path works.
Qwen3-VL 235B-A22B versus 30B-A3B
The complete official Q4 paths are 142.906GB for 235B-A22B and 19.269GB for 30B-A3B with Q8_0 projectors. Their cache payloads are 188KiB and 96KiB per token respectively. These facts describe capacity, not a universal performance or quality ranking.
Compare the Qwen3-VL 30B-A3B guide, dense Qwen3-VL family hub, MoE VRAM explainer and 48GB–96GB capacity guide.
Reproducible flagship fit test
- Record repository, revision, Instruct/Thinking variant, every language shard, projector and checksum.
- Record runtime/version, MoE kernel and image/video support.
- Record every device, memory capacity, topology, partitioning mode and offload target.
- Record model, projector and cache precision plus per-device placement.
- Fix image dimensions/count, video sampling, text allocation, output target and concurrency.
- Measure accelerator and system memory per device before load, after load, after vision encoding and at peak generation.
- Capture preprocessing time, time to first token, throughput, communication/fallback behavior and errors.
Use the VRAM calculator, multi-GPU guide, workload test and benchmark submission.
Continue planning
Browse all guides and tools. Validate per-device peak memory rather than summing labels on a parts list.
MoE family comparison: the Qwen3-VL MoE VRAM hub contrasts this flagship with the 30B-A3B path using exact file totals and cache math.
Sources and limits
- Qwen official Qwen3-VL 235B-A22B repository
- Qwen official configuration
- Qwen official BF16 index
- Qwen official 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 runtime, topology, MoE kernel, image/video workload, variant or quantization. File size is not peak resident memory.
Newer hybrid MoE comparison: Qwen3.5-122B-A10B covers official GPTQ, FP8 and BF16 paths with 24KiB-per-token full-attention KV geometry.
Newer flagship comparison: use the Qwen3.5-397B-A17B VRAM guide for exact official 235.708GB–806.796GB checkpoints and hybrid-attention planning.