Qwen3.5 VRAM Requirements: 0.8B to 397B-A17B Compared

Newer generation: Qwen3.8 is documented separately. See Qwen3.8-27B VRAM requirements for exact community GGUF and official checkpoint sizes; do not assume the Qwen3.5 table applies unchanged.

Short answer: Qwen3.5 spans edge devices through multi-accelerator systems. Practical starting points range from 2GB for quantized 0.8B, through 4GB for 2B, 6GB for 4B, 12GB for 9B, 24GB for dense 27B Q4, 32GB for 35B-A3B 4-bit, 96GB for 122B-A10B 4-bit, and 320GB for 397B-A17B 4-bit. These are workload-planning tiers, not speed or quality rankings.

Qwen3.5 Q4, Q8 and BF16 file totals

Model	Community Q4 + F16 projector	Community Q8 + F16 projector	Official BF16
0.8B	0.738GB	1.017GB	1.747GB
2B	1.949GB	2.680GB	4.548GB
4B	3.413GB	5.155GB	9.320GB
9B	6.599GB	10.446GB	19.306GB
27B	17.668GB	29.523GB	55.563GB
35B-A3B	22.915GB	37.802GB	71.904GB
122B-A10B	77.446GB	130.781GB	250.173GB
397B-A17B	245.012GB	422.426GB	806.796GB

Practical starting tiers by model

Model	Q4 starting point	Q8 starting point	Official BF16 path
0.8B	1GB constrained; 2GB practical	2GB practical	2GB tight; 4GB safer
2B	2GB very tight; 4GB practical	4GB practical	6GB bounded; 8GB practical
4B	4GB tight; 6GB practical	6GB tight; 8GB practical	12GB bounded; 16GB practical
9B	8GB tight; 12GB practical	12GB tight; 16GB practical	24GB bounded; 32GB practical
27B	24GB practical	32GB tight; 48GB practical	64GB bounded; 80GB safer
35B-A3B	24GB very tight; 32GB practical	48GB practical	80GB bounded; 96GB practical
122B-A10B	80GB very tight; 96GB practical	128GB does not hold complete community files; 160GB practical	256GB very tight; 320GB practical
397B-A17B	256GB very tight; 320GB practical	480GB tight; 512GB practical	800GB does not hold files; 1TB practical

Dense models versus Qwen3.5 MoE

The 0.8B, 2B, 4B, 9B and 27B releases are dense language backbones. The 35B-A3B, 122B-A10B and 397B-A17B releases are sparse MoE models. Their A3B, A10B and A17B labels describe routed computation, not stored-weight memory. All expert weights still require storage or placement: the official configurations document 256 experts with eight routed for 35B and 122B, then 512 experts with ten routed for 397B.

Hybrid-attention cache geometry differs by size

Models	Language layers	Full-attention layers	FP16/BF16 full-attention KV only	32K payload
0.8B and 2B	24	6	12KiB/token	0.375GiB
4B and 9B	32	8	32KiB/token	1GiB
27B	64	16	64KiB/token	2GiB
35B-A3B	40	10	20KiB/token	0.625GiB
122B-A10B	48	12	24KiB/token	0.75GiB
397B-A17B	60	15	30KiB/token	0.9375GiB

Choose by workload, not parameter count alone

• Edge prototype: start with 0.8B or 2B, then validate output quality for the actual task.
• Compact multimodal work: 4B and 9B add capacity while remaining accessible to common 8GB–16GB hardware through quantization.
• Higher-capability dense path: 27B starts at practical 24GB Q4 but needs substantially more memory for Q8 or BF16.
• Sparse MoE paths: 35B-A3B, 122B-A10B and 397B-A17B reduce active computation relative to total parameters but do not turn their stored weights into 3B, 10B or 17B models.
• Multi-device scale: for 122B and 397B, validate per-device placement, parallelism and interconnect topology; summed capacity is not automatic pooling.
• Images and video: include the matching projector, verify the actual visual path and fix preprocessing parameters before comparing memory.
• Long context: native 256K capability is not a promise that your hardware can allocate the chosen workload.

Validate before buying

1. Choose the exact model, repository, revision and quantization.
2. Record every language/projector file and checksum; do not mix first-party and community provenance.
3. Verify runtime support for Qwen3.5, hybrid attention, MoE when applicable and image/video input.
4. Fix context allocation, image/video workload, output target and concurrency.
5. Record weight, projector, full-attention KV and linear-state precision plus device placement/offload.
6. Measure accelerator and system memory before load, after load, after vision encoding and at peak generation.
7. Capture preprocessing time, time to first token, throughput, fallback and errors.

Use the multimodal VRAM hub, VRAM calculator, workload test, benchmark submission and all guides and tools.

Sources and limits

• Qwen official Qwen3.5 collection
• Qwen3.5 official code repository
• Each detailed guide links its model-specific official repository, configuration and checkpoint index plus the separately labeled community conversion.
• HighVRAM methodology

Reviewed October 25, 2026. File totals are direct repository sums documented in the eight detailed guides. HighVRAM has not independently benchmarked every model, runtime, kernel, conversion, quantization or multimodal workload. Fit does not establish speed, output quality or value.

Large-model hardware: compare workstation, unified-memory and datacenter paths for Qwen3.5-122B and 397B from manufacturer-published capacities.