Short answer: treat 4GB as a tight Q4 boundary, 6GB as the practical Q4 tier, 8GB as the practical Q8 tier, and 12GB/16GB as the bounded/practical tiers for Qwen’s official 9.320GB BF16 checkpoint. Qwen3.5-4B is natively multimodal, so a GGUF image/video deployment needs the matching projector as well as the language file.
Source boundary: Qwen publishes the BF16 checkpoint and configuration. The Q4/Q8 GGUF and projector files below are an Unsloth community conversion, not first-party Qwen quantizations. Verify repository revision, checksums and runtime compatibility.
Qwen3.5-4B files and practical tiers
| Deployment path | Displayed files | Combined size | Planning tier |
|---|---|---|---|
| Community Q4_K_M + F16 projector | 2.741GB + 0.672GB | 3.413GB | 4GB tight; 6GB practical |
| Community Q8_0 + F16 projector | 4.482GB + 0.672GB | 5.155GB | 6GB tight; 8GB practical |
| Official Qwen BF16 safetensors | Two shards | 9.320GB | 12GB bounded; 16GB practical |
Why 4GB Q4 and 6GB Q8 are tight
The community Q4 language file plus F16 projector occupies 3.413GB, leaving about 0.6GB of nominal space on a 4GB device before cache, hybrid-attention state, vision processing, compute buffers, allocator behavior or display use. The Q8 pair occupies 5.155GB, leaving about 0.8GB on a 6GB device. Offload can allow a workload to launch, but that is not a whole-model GPU fit and does not guarantee useful latency.
Hybrid attention: calculate only the justified component
Qwen’s official model card and configuration describe 32 language layers arranged as eight repetitions of three Gated DeltaNet linear-attention layers followed by one full-attention layer. That yields 24 linear-attention layers and eight full-attention layers. The full-attention blocks use four KV heads with a 256-element head dimension.
With FP16/BF16 keys and values, the sequence-linear full-attention component is 2 × 8 × 4 × 256 × 2 = 32,768 bytes, or 32KiB per allocated token. Do not apply a conventional KV-cache formula to all 32 layers: the linear-attention layers maintain implementation-dependent recurrent and convolution state.
| Allocated tokens | Full-attention KV payload only |
|---|---|
| 4K | 0.125GiB |
| 8K | 0.25GiB |
| 32K | 1GiB |
| 64K | 2GiB |
| 256K native maximum | 8GiB |
Compact-device safeguards
- Projector: include the conversion’s 0.672GB F16 projector for image/video inference.
- Vision: test the actual image/video path; text-only output does not prove projector compatibility.
- Runtime: verify Qwen3.5, Gated DeltaNet and multimodal support in the exact release.
- MTP: record whether multi-token-prediction components are loaded or used.
- Memory type: unified memory and system RAM are not dedicated VRAM; report placement and offload.
- Workload: fix context allocation, image dimensions/count, video sampling, output length and concurrency.
Qwen3.5-4B versus Qwen3.5-9B
The Qwen3.5-9B guide covers a larger dense sibling whose community Q4-plus-projector pair is 6.599GB and official BF16 checkpoint is 19.306GB. Both use a 32-layer 3:1 hybrid attention pattern, but equal cache geometry does not mean equal total memory, speed, quality or visual accuracy.
Reproducible Qwen3.5-4B fit test
- Record the first-party repository, conversion repository, revisions, every file and checksum.
- Record runtime/version and verify 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 text allocation, image/video workload, 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.
Continue with the Qwen3-VL 4B comparison, 3B/4B sizing guide, multimodal VRAM hub, 8GB–16GB capacity guide, VRAM calculator, workload test, benchmark submission or all guides and tools.
Sources and limits
- Qwen official Qwen3.5-4B repository
- Qwen official configuration
- Qwen official BF16 index
- Unsloth community GGUF conversion
- HighVRAM methodology
Reviewed October 25, 2026. Repository file sizes were summed directly. Qwen’s BF16 checkpoint and configuration 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.
Smaller sibling: the Qwen3.5-2B VRAM guide maps a 1.949GB community Q4-plus-projector pair and 4.548GB official BF16 checkpoint to 2GB–8GB 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.
Choosing hardware for this 8GB planning tier? Use the RTX 4060 versus RTX 5060 buyer guide to keep the same model-fit ceiling separate from architecture, software support, power and measured performance.
Testing this model on an entry Ada card? Use the RTX 4060 8GB local-AI guide to connect these exact 4B file totals to software, power, display-use and upgrade checks.
Testing this model on an entry Blackwell card? Use the RTX 5060 8GB local-AI guide to connect these exact 4B file totals to compute-capability, power, display-use and upgrade checks.