Short answer: plan 32GB for a practical 4-bit deployment, 48GB for a practical Q8 deployment, and 80GB/96GB for the official 71.904GB BF16 checkpoint. A 24GB card is a boundary case, not a comfortable default. Qwen3.5-35B-A3B is a natively multimodal mixture-of-experts model, so artifact completeness, vision support and runtime support all matter.
Do not read A3B as a 3B storage requirement. Qwen documents 35B total parameters and 3B activated per token. The full expert set still has to be stored or placed across accelerator and system memory.
Qwen3.5-35B-A3B files and practical tiers
| Deployment path | Displayed files | Total | Planning tier |
|---|---|---|---|
| Official Qwen GPTQ-Int4 safetensors | 14 shards | 24.420GB | 24GB does not hold the files; 32GB practical |
| Community Q4_K_M + F16 projector | 22.016GB + 0.899GB | 22.915GB | 24GB very tight; 32GB practical |
| Community Q8_0 + F16 projector | 36.903GB + 0.899GB | 37.802GB | 48GB practical; 64GB roomier |
| Official Qwen BF16 safetensors | 14 shards | 71.904GB | 80GB bounded; 96GB practical |
Why 24GB is only a boundary case
The official GPTQ-Int4 checkpoint totals 24.420GB, so its files alone exceed 24GB before cache, hybrid-attention state, vision processing, compute buffers or allocator overhead. The community Q4_K_M plus F16 projector totals 22.915GB, leaving roughly 1.1GB of nominal space on a 24GB device before those runtime allocations. Offload may make a workload launch, but that is not a whole-model 24GB GPU fit and does not establish usable speed.
Hybrid attention: calculate only what is justified
Qwen’s official model card and configuration describe 40 language layers arranged as ten repetitions of three Gated DeltaNet linear-attention layers followed by one full-attention layer. That means 30 linear-attention layers and ten full-attention layers. The full-attention blocks use two KV heads with a 256-element head dimension.
For FP16/BF16 keys and values, the sequence-linear full-attention component is 2 × 10 × 2 × 256 × 2 = 20,480 bytes, or 20KiB per allocated token. Do not apply a conventional KV-cache formula to all 40 layers: the linear-attention blocks maintain implementation-dependent recurrent and convolution state.
| Allocated tokens | Full-attention KV payload only |
|---|---|
| 8K | 0.15625GiB |
| 32K | 0.625GiB |
| 64K | 1.25GiB |
| 128K | 2.5GiB |
| 256K native maximum | 5GiB |
MoE and multimodal safeguards
- Experts: Qwen documents 256 experts, eight routed experts and one shared expert; active computation is not stored-weight memory.
- Vision: verify image and video inference, not only text generation; GGUF deployments need the matching projector.
- Runtime: verify the exact release supports Qwen3.5 MoE, Gated DeltaNet, multimodal input and the chosen quantization.
- MTP: Qwen documents multi-token-prediction training and runtime options; record whether it is loaded or used.
- Placement: record GPU layers, expert placement, offload, tensor/expert parallelism and system RAM.
- Workload: fix text allocation, image dimensions/count, video sampling, output target and concurrency before comparing peaks.
Qwen3.5-35B-A3B versus Qwen3.5-27B
The 35B-A3B model is sparse: all 35B parameters require storage, while about 3B are activated per token. The Qwen3.5-27B guide covers a dense 27B alternative whose community Q4-plus-projector pair is 17.668GB and whose official BF16 checkpoint is 55.563GB. These architecture and file-size differences do not establish a universal speed, quality or visual-accuracy winner.
Reproducible fit test
- Record repository, revision, every weight/projector file and checksums.
- Record runtime and version; verify Qwen3.5 MoE, hybrid attention and image/video support.
- Record weight, projector, full-attention KV and linear-state precision.
- Record accelerator placement, expert/tensor parallelism, offload and MTP behavior.
- Fix context allocation, image/video workload, output length and concurrency.
- Measure accelerator and system memory before load, after load, after vision encoding and at peak generation.
- Capture preprocessing time, time to first token, throughput, fallback and errors.
Continue with the 27B/32B sizing guide, Qwen3-VL 30B-A3B comparison, multimodal VRAM hub, 24GB–48GB capacity guide, VRAM calculator, workload test, benchmark submission or all guides and tools.
Sources and limits
- Qwen official model card
- Qwen official configuration
- Qwen official BF16 index
- Qwen official GPTQ-Int4 repository
- Unsloth community GGUF conversion
- HighVRAM methodology
Reviewed October 25, 2026. Repository file sizes were summed directly. Qwen’s BF16 and GPTQ artifacts are first-party; GGUF/projector sizes are from the separately labeled Unsloth conversion. HighVRAM has not independently benchmarked every runtime, kernel, conversion, quantization or multimodal workload. File size is not peak resident memory.
32GB-versus-96GB hardware path: compare RTX 5090 with RTX PRO 6000 Blackwell for this model’s exact GPTQ, Q4, Q8 and BF16 file boundaries, plus ECC, MIG and support differences.
32GB versus shared-memory path: compare the Mac Studio M5 Ultra and RTX 5090 using the exact 35B Qwen artifact boundaries before choosing the faster CUDA route or the larger unified-memory route.
32GB versus 48GB hardware paths: compare RTX 5090 with either the RTX A6000 or RTX 6000 Ada using this guide’s exact 35B GPTQ, Q4, Q8-plus-projector and BF16 totals; keep Ampere/Ada generation, power, ECC, support and used-card differences separate. For the 24GB-to-32GB generation decision, use the RTX 3090 vs 4090 vs 5090 comparison.
BF16 upper-capacity path: compare the exact 71.904GB checkpoint across the 48GB, 64GB, 80GB and 96GB capacity tiers before treating an 80GB file fit as safe runtime headroom.
80GB, 94GB or 96GB platform path: compare RTX PRO 6000 Blackwell with H100 using the 35B BF16 file margin while keeping workstation cooling and Hopper data-center integration distinct.
Compare the family: use the Qwen3.5 VRAM hub to compare complete file totals, practical tiers and dense-versus-MoE storage behavior from 0.8B through 35B-A3B.
Larger sparse sibling: the Qwen3.5-122B-A10B guide covers official 78.864GB GPTQ, 127.163GB FP8 and 250.173GB BF16 checkpoints.
Track matching hardware
Save a private browser-only hardware watchlist or choose a focused RSS feed. No account or email address is required; the watchlist stays in this browser and does not send price alerts.
48GB versus 80GB/94GB hardware path: compare RTX 6000 Ada with NVIDIA H100 using this model’s exact Q8 and BF16 file margins while keeping workstation and data-center deployment distinct.
48GB Ampere versus 80GB/94GB Hopper: compare RTX A6000 with NVIDIA H100 using this model’s exact Q8 and BF16 file margins while keeping used-workstation and data-center deployment distinct.
Same-capacity 48GB hardware fork: compare RTX A6000 with RTX 6000 Ada. The 10.198GB Q8-plus-projector file margin is identical, so measure architecture/software effects separately.
Ampere 40GB, 48GB or 80GB path: compare RTX A6000 with A100 using this guide’s 37.802GB Q8-plus-projector and 71.904GB BF16 totals.
Ada 48GB versus Ampere 40GB/80GB path: compare RTX 6000 Ada with A100 using this guide’s 37.802GB Q8-plus-projector and 71.904GB BF16 totals.
Server 40GB, 48GB or 80GB path: compare L40S with A100 using this guide’s 37.802GB Q8-plus-projector and 71.904GB BF16 totals.
Ada 48GB versus Hopper 80GB/94GB server path: compare L40S with H100 using this guide’s 37.802GB Q8-plus-projector and 71.904GB BF16 totals.
Ampere server 40GB, 48GB or 80GB path: compare A40 with A100 using this guide’s 37.802GB Q8-plus-projector and 71.904GB BF16 totals.
Applying these files to a server upgrade? The A40 versus H100 comparison maps the exact 37.802GB and 71.904GB totals across 48GB, 80GB and 94GB devices with runtime-headroom caveats.
Applying the 35B files to two 48GB NVIDIA paths? Compare A40 with RTX 6000 Ada; both leave the same file-only margins, but cooling, connector, NVLink and software differ.
Applying the 35B files to 32GB and 48GB? Compare RTX 5090 with A40, including the 32GB 4-bit margins and the 48GB Q8 capacity change.
Choosing hardware for this model? Compare RTX 5090 32GB with A100 40GB and 80GB using exact Q4, Q8, GPTQ and BF16 file margins.
Choosing hardware for this model? Compare RTX 5090 32GB with H100 80GB and 94GB using exact Q8 and BF16 file margins.
Choosing hardware for this model? Compare RTX 5090 32GB with H200 141GB using exact BF16 file margins and platform safeguards.
Hardware route: compare the listed 35B artifacts on RTX 5090 and DGX Spark, with exact file margins and Arm-platform caveats.
Compact-system route: compare the listed 35B artifacts in DGX Spark vs Mac Studio, with nominal shared-memory margins clearly labeled.
Hardware route: compare the listed 35B artifacts in DGX Spark vs RTX PRO 6000 with exact file-only margins.
DGX Spark vs H100: Map the listed 35B artifact to DGX Spark and H100 in the exact capacity and platform comparison.
DGX Spark vs H200: Map the exact 35B artifact to DGX Spark and H200 in the exact 128GB-versus-141GB comparison.
For complete-system versus accelerator tradeoffs at this model size, compare DGX Spark vs A100.
The direct 24GB-versus-128GB hardware decision is covered in RTX 4090 vs DGX Spark, including the unsafe 22.915GB Q4 boundary.
The used 24GB versus 128GB system decision is covered in RTX 3090 vs DGX Spark, including the unsafe 22.915GB Q4 boundary.
Choosing a 32GB device for this model? Compare RTX 5090 and RTX 5000 Ada using the same exact Q4, GPTQ and BF16 boundaries plus power, ECC and workstation-support tradeoffs.