Short answer: plan 24GB for a practical Q4 deployment, 32GB as a bounded Q8 target, 48GB as the practical Q8 tier, and 64GB/80GB for the official 55.563GB BF16 checkpoint. Qwen3.5-27B is natively multimodal, so a GGUF deployment also needs a vision projector.
Source boundary: Qwen publishes the BF16 checkpoint and configuration. The Q4/Q8 GGUF and projector files below are a community conversion by Unsloth, not first-party Qwen artifacts. Revisions, checksums and runtime compatibility must be verified independently.
Qwen3.5-27B files and practical tiers
| Deployment path | Displayed files | Combined file size | Planning tier |
|---|---|---|---|
| Community Q4_K_M + F16 projector | 16.741GB + 0.928GB | 17.668GB | 24GB practical; 32GB roomier |
| Community Q8_0 + F16 projector | 28.596GB + 0.928GB | 29.523GB | 32GB tight; 48GB practical |
| Official Qwen BF16 safetensors | Checkpoint index totaling 55.563GB | 55.563GB | 64GB bounded; 80GB safer |
Qwen3.5 uses hybrid attention
The official configuration declares 64 language layers in a repeating pattern of three linear-attention layers followed by one full-attention layer. That yields 48 Gated DeltaNet-style linear layers and 16 full-attention layers. Do not apply a conventional all-layers KV-cache formula to the entire model.
The sequence-linear full-attention component uses 16 layers, four KV heads and a 256-element head dimension. With FP16/BF16 keys and values, it contributes 2 × 16 × 4 × 256 × 2 = 65,536 bytes, or 64KiB per allocated token. The linear-attention layers also maintain recurrent/convolution state whose layout and precision depend on the implementation.
| Allocated tokens | Full-attention KV payload only |
|---|---|
| 8K | 0.5GiB |
| 32K | 2GiB |
| 64K | 4GiB |
| 128K | 8GiB |
| 256K architectural maximum | 16GiB |
Why 32GB Q8 is tight
The community Q8 language file plus F16 projector already occupies 29.523GB. A 32GB device has little nominal room for full-attention KV payload, linear-attention state, vision processing, allocator behavior, compute buffers or display use. Forty-eight gigabytes is the practical Q8 tier for a bounded workload.
The official BF16 checkpoint is 55.563GB before runtime allocations. Sixty-four gigabytes is therefore bounded; 80GB provides safer headroom but still requires context, images/video and concurrency to be specified.
Multimodal and runtime safeguards
- Projector: include the conversion’s 0.928GB F16 projector for image/video inference.
- Vision workload: image dimensions/count and video frame sampling affect tokens and intermediate state.
- Hybrid kernels: verify Gated DeltaNet/linear-attention support in the exact runtime version.
- MTP: the configuration declares one multi-token-prediction layer; record whether the runtime loads or uses it.
- Cache/state precision: record full-attention KV precision and any linear-state precision separately.
- Offload and concurrency: device placement and parallel requests change peak accelerator and system memory.
Qwen3.5-27B versus Qwen3-VL 32B
Qwen3.5-27B is a newer natively multimodal hybrid-attention architecture; Qwen3-VL 32B uses conventional full attention in all 64 language layers. Their artifact formats and cache behavior are not interchangeable. The community Qwen3.5 Q4 pair is 17.668GB, while Qwen’s official Qwen3-VL 32B Q4 pair is 20.535GB. Those sizes do not establish a universal speed, visual-accuracy or quality winner.
Compare the Qwen3-VL 32B guide, Qwen3-VL dense family, 27B/32B sizing guide and multimodal VRAM hub.
Reproducible Qwen3.5-27B fit test
- Record the first-party base repository, conversion repository, revisions, every file and checksum.
- Record runtime/version and verify Qwen3.5 hybrid-attention plus image/video support.
- Record weight, projector, full-attention KV and linear-state precision.
- Record device placement, offload, MTP behavior and kernel/fallback path.
- Fix image dimensions/count, video sampling, text allocation, output target 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.
Use the 24GB–48GB capacity guide, VRAM calculator, workload test and benchmark submission.
Continue planning
Browse all guides and tools. Treat community conversion evidence and first-party checkpoint evidence as separate provenance classes.
Sources and limits
- Qwen official Qwen3.5-27B repository
- Qwen official configuration
- Qwen official BF16 index
- Unsloth community GGUF conversion
- Qwen3.5 official code repository
- HighVRAM methodology
Reviewed October 25, 2026. Qwen’s BF16 index and configuration are first-party. GGUF and projector sizes are from the separately labeled Unsloth conversion. HighVRAM has not independently benchmarked every runtime, conversion, kernel, 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 Q4, Q8 and BF16 file boundaries, then measure the intended multimodal runtime.
32GB versus 48GB hardware paths: compare RTX 5090 with either the RTX A6000 or RTX 6000 Ada using this guide’s exact 27B 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 55.563GB checkpoint across the 48GB, 64GB, 80GB and 96GB capacity tiers before treating a 64GB file fit as safe runtime headroom.
Choosing hardware for this model? Use the 27B GPU buyer guide to map the exact Q4, Q8 and BF16 boundaries to 24GB, 32GB, 48GB and 64GB–80GB paths.
Sparse Qwen3.5 alternative: the Qwen3.5-35B-A3B VRAM guide covers official GPTQ/BF16 artifacts, community GGUF paths and the active-versus-stored MoE distinction.
Mainstream sibling: the Qwen3.5-9B VRAM guide maps a 6.599GB community Q4-plus-projector pair and 19.306GB official BF16 checkpoint to 8GB–32GB hardware.
Compare the family: use the Qwen3.5 VRAM hub to compare complete file totals, practical tiers and hybrid-cache geometry from 0.8B through 35B-A3B.
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.
Same-capacity 48GB hardware fork: compare RTX A6000 with RTX 6000 Ada. This model has identical file margins on both, so measure architecture/software effects separately.
Applying these files to two 48GB NVIDIA paths? The A40 versus RTX 6000 Ada comparison preserves identical 27B capacity arithmetic while separating architecture and system requirements.
Applying these files to 32GB and 48GB NVIDIA paths? The RTX 5090 versus A40 comparison maps the exact 27B Q4/Q8/BF16 totals to both devices with runtime safeguards.
Choosing hardware for this model? Compare RTX 5090 32GB with A100 40GB and 80GB using the artifact sizes on this page and explicit runtime headroom.
Choosing hardware for this model? Compare RTX 5090 32GB with H100 80GB and 94GB using the artifact sizes on this page and explicit runtime headroom.
Choosing hardware for this model? Compare RTX 5090 32GB with H200 141GB using the artifact sizes on this page and explicit runtime headroom.
Hardware route: compare the listed 27B artifacts on RTX 5090 and DGX Spark, with exact file margins and unified-memory caveats.
For this model’s exact 17.668GB Q4, 29.523GB Q8 and 55.563GB BF16 paths, compare RTX 4090 24GB with DGX Spark 128GB.
For this model’s 17.668GB Q4, 29.523GB Q8 and 55.563GB BF16 paths, compare RTX 3090 24GB with DGX Spark 128GB.