Short answer: treat 8GB as a tight Q4 boundary, 12GB as the practical Q4 tier, 16GB as the practical Q8 tier, and 24GB/32GB as the bounded/practical tiers for Qwen’s official 19.306GB BF16 checkpoint. Qwen3.5-9B is natively multimodal, so a GGUF vision 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 revisions, checksums and runtime compatibility independently.
Qwen3.5-9B files and practical tiers
| Deployment path | Displayed files | Combined size | Planning tier |
|---|---|---|---|
| Community Q4_K_M + F16 projector | 5.681GB + 0.918GB | 6.599GB | 8GB tight; 12GB practical |
| Community Q8_0 + F16 projector | 9.528GB + 0.918GB | 10.446GB | 12GB tight; 16GB practical |
| Official Qwen BF16 safetensors | Four shards | 19.306GB | 24GB bounded; 32GB practical |
Why 8GB Q4 and 12GB Q8 are tight
The community Q4 language file plus F16 projector occupies 6.599GB, leaving roughly 1.4GB of nominal space on an 8GB device before cache, hybrid-attention state, vision processing, compute buffers, allocator behavior or display use. The Q8 pair occupies 10.446GB, leaving roughly 1.6GB on a 12GB device. Both can be valid bounded experiments, but neither file-size remainder is a guarantee that a chosen workload stays fully resident.
Hybrid attention: calculate only the full-attention 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 |
|---|---|
| 8K | 0.25GiB |
| 32K | 1GiB |
| 64K | 2GiB |
| 128K | 4GiB |
| 256K native maximum | 8GiB |
Multimodal and runtime safeguards
- Projector: include the conversion’s 0.918GB 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 and linear-attention support in the exact runtime release.
- MTP: Qwen documents multi-token-prediction training; record whether the runtime loads or uses it.
- Precision: record weights, projector, full-attention KV and linear-state precision separately.
- Offload: system-memory offload can enable launch but is not dedicated VRAM and can change latency.
Qwen3.5-9B versus Qwen3-VL 8B
Qwen3.5-9B uses a hybrid 3:1 linear/full-attention stack. The Qwen3-VL 8B guide covers a different conventional full-attention architecture and first-party Qwen GGUF artifacts. Do not transfer cache formulas or artifact provenance between them. Their file sizes do not establish a universal speed, quality or visual-accuracy winner.
Reproducible Qwen3.5-9B 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.
Continue with the Qwen3.5-27B guide, 8B 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 Qwen3.5-9B 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.
Compact sibling: the Qwen3.5-4B VRAM guide maps a 3.413GB community Q4-plus-projector pair and 9.320GB official BF16 checkpoint to 4GB–16GB 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 a 12GB GPU for this model? Compare RTX 3060 and RTX 5070 using the same Q4/Q8/projector boundaries plus runtime, power and complete-system checks.
Testing this model on a current 12GB Blackwell card? Use the RTX 5070 12GB local-AI guide to connect these exact 9B totals to capability, power, display-use and upgrade checks.
Choosing between two Blackwell cards for this model? The RTX 5070 12GB vs RTX 5060 Ti 16GB guide shows why the listed Q8 pair has unsafe 1.554GB versus more workable 5.554GB file-only margin.