How much VRAM to run Qwen3.6 35B-A3B?

About 25 GB atQ4_K_M with an 8K context — fits a RTX 5090. Full breakdown below, or check your exact hardware.

Qwen3.6 35B-A3B VRAM by quantisation

QuantisationWeightsTotal (8K ctx)Fits on
Q4_K_M21.4 GB24.7 GBRTX 5090, Radeon AI PRO R9700
Q5_K_M24.9 GB28.6 GBRTX 5090, Radeon AI PRO R9700
Q6_K28.7 GB32.8 GBApple Silicon 48GB unified, RTX 6000 Ada
Q8_037.3 GB42.2 GBRTX 6000 Ada, Apple Silicon 64GB unified
FP16 / BF1670.0 GB78.2 GBA100 80GB, RTX PRO 6000 Blackwell

Check your hardware

About Qwen3.6 35B-A3B

Qwen3.6 35B-A3B led Alibaba's Qwen3.6 series into the open: a mixture-of-experts release whose active-parameter count per token is a small slice of the total, drawn a few experts at a time from a large routed pool plus one shared expert. Its card leans on agentic coding — frontend workflows, repository-level reasoning — alongside maths, tool calling and vision-language work, all under Apache 2.0.

The design goal reads clearly from the numbers: dense-mid-size quality at small-model generation speed, with the generation's hybrid linear-attention layout carrying the long-context load.

Qwen3.6 35B-A3B is Alibaba's 35B-parameter model released in April 2026, with a 256K-token context window. It is a mixture-of-experts model: all 35B parameters must sit in memory, but only ~3B 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 0.2 GB at an 8K context, 2.7 GB at 128K, and 5.4 GB at the full 256K window (FP16 cache).

MoE memory rules apply — every expert stays resident, so plan from the full parameter count above, which lands quality quantisations on high-VRAM consumer cards and up — but the payoff is generation speed closer to a model a tenth the size, because so few parameters activate per token. Layer on the hybrid attention design, whose linear blocks hold constant-size state, and you get an unusual profile: a model that is briskly interactive and cheap at long context once you clear its resident-weight threshold. Below that threshold, the dense sibling at lower quantisation is usually the better compromise.

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 Qwen3.6 35B-A3B need?

At Q4_K_M with an 8K context, Qwen3.6 35B-A3B needs about 25 GB (weights 21 GB + KV cache + overhead). The smallest common hardware that fits is a RTX 5090.

Can an RTX 4090 (24GB) run Qwen3.6 35B-A3B?

Not fully in VRAM. Qwen3.6 35B-A3B needs about 25 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 Qwen3.6 35B-A3B?

Yes — Apple Silicon with 36 GB of unified memory or more (macOS lets the GPU use ~75% of it, ~27 GB) runs Qwen3.6 35B-A3B at Q4_K_M.

How much VRAM does Qwen3.6 35B-A3B need at its full 256K context?

The KV cache grows with context: about 0.2 GB at 8K tokens versus 5.4 GB at the full 256K window (FP16 cache). Add that to the weights (21 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: Gemma 4 31B (31B) · Qwen3 Coder 30B-A3B (30.5B) · Qwen3.6 27B (27B) · Gemma 4 26B-A4B (26B) · GPT-OSS 20B (21B)

Last updated 2026-09-03. Architecture figures from the model's published config.json; see the methodology.