About Qwen3.5 397B-A17B
Qwen3.5 397B-A17B is the open flagship of Alibaba's Qwen3.5 generation: a large mixture-of-experts system with a compact active set per token, the generation's hybrid linear-attention architecture, and the full agentic toolkit — thinking modes, tool calling, vision-language input — released under Apache 2.0 rather than held behind an API.
It represents the pattern the open Chinese labs have made standard: frontier-scale weights published outright, with the expectation that serious users run them on shared infrastructure and everyone else meets them through hosted providers.
Qwen3.5 397B-A17B is Alibaba's 397B-parameter model released in February 2026, with a 256K-token context window. It is a mixture-of-experts model: all 397B parameters must sit in memory, but only ~17B are active per token, which is what makes it fast for its size. It uses a hybrid-attention design — only a fraction of its layers cache the full context, so long conversations cost far less VRAM than a classic dense model: its KV cache is about 0.3 GB at an 8K context, 4.0 GB at 128K, and 8.1 GB at the full 256K window (FP16 cache).
The totals above are datacenter numbers — the full expert pool dwarfs any single consumer or workstation card at any quantisation, so realistic homes are multi-accelerator servers and the very largest unified-memory configurations at aggressive quantisation. Two properties keep it surprisingly tractable at that scale: the small active set makes per-token compute light for the size, and the linear-attention-heavy layout means even its extended context adds little cache on top of the weights. Treat this page as capacity planning; for a desktop-class taste of the generation, its mid-size siblings exist for exactly that reason.
For most people Q4_K_M is the sweet spot — the most popular quality/size trade-off — while Q8 is near-lossless if you have the memory. Totals above include the KV cache and a realistic framework overhead, so they are what you should expect to see in practice rather than just the download size. Weight sizes are calibrated against real GGUF files — see the methodology.
Frequently asked questions
How much VRAM does Qwen3.5 397B-A17B need?
At Q4_K_M with an 8K context, Qwen3.5 397B-A17B needs about 268 GB (weights 242 GB + KV cache + overhead). The smallest common hardware that fits is a Apple Silicon 512GB unified.
Can an RTX 4090 (24GB) run Qwen3.5 397B-A17B?
Not fully in VRAM. Qwen3.5 397B-A17B needs about 268 GB even at Q4_K_M, so a 24 GB card would have to offload layers to system RAM at a large speed penalty.
Can a Mac run Qwen3.5 397B-A17B?
Yes — Apple Silicon with 512 GB of unified memory or more (macOS lets the GPU use ~75% of it, ~384 GB) runs Qwen3.5 397B-A17B at Q4_K_M.
How much VRAM does Qwen3.5 397B-A17B need at its full 256K context?
The KV cache grows with context: about 0.3 GB at 8K tokens versus 8.1 GB at the full 256K window (FP16 cache). Add that to the weights (242 GB at Q4_K_M) plus overhead — long contexts can cost more than a whole quantisation step.
Related
VRAM calculator for any model ·Token counter
Similar-size local models: Llama 4 Maverick (400B) · DeepSeek V4 Flash (284B) · Mistral Large 3 (675B) · GLM-5.2 (753B) · Mistral Small 4 (119B)
Last updated 2026-09-03. Architecture figures from the model's published config.json; see the methodology.