This post contains affiliate links. If you buy through them, Aliteq may earn a commission — at no extra cost to you. Prices verified at publish time.
Share
Here's a fact worth knowing before you download anything: when you type ollama run qwen3.5, the model you get is this one, the 9B. On ollama.com/library, 9b is the default tag the bare name resolves to. So Qwen3.5 9B is quietly one of the most-run local models simply because it's what the obvious command pulls.
What Qwen3.5 9B actually is
From the Qwen3.5 9B model card (read 26 Sep 2026): 9B parameters, a 262,144-token (256K) context natively, extensible further, released under the Apache-2.0 licence — which means it's genuinely free to use, including commercially. It's a dense model.
Its config.json shows an efficiency trick that keeps memory low: most layers (24 of 32) use linear attention, with 8 full-attention layers. Linear-attention layers carry a tiny fixed state instead of a growing KV cache, so long contexts cost far less VRAM than the raw parameter count would suggest.
That matters more than it sounds. A 9B model is already small enough for most cards, but the linear-attention layers mean the 256K context doesn't punish you: filling it doesn't blow up your VRAM the way it would on an older 9B. Combined with Apache-2.0 licensing, that makes Qwen3.5 9B a genuinely low-friction default — it's cheap on memory, free to use, and it's the model the obvious command already pulls, which is a large part of why it's so widely run.
At ~6.7 GB, Q4_K_M fits an 8GB card. A 12GB card runs Q8 (near-full quality) with room for a long context — the comfortable choice for a 9B.
If the quant labels are new to you, here's the short version. Q4_K_M compresses each weight to roughly four bits, shrinking the model to about a quarter of its full size while keeping quality close to the original — it's the default most people run, and the one Ollama pulls unless you ask otherwise. Q6_K and Q8_0 spend more VRAM to claw back the last few percent of quality, and BF16 is the uncompressed model you'd only run on a big or rented card. For a 9B, the two combinations worth remembering are Q4_K_M on an 8GB card and Q8 on a 12GB card.
The 256K context is more than most local work needs — chat and coding rarely run past a few thousand tokens — but it's there when you paste in a long document, and, thanks to those linear-attention layers, using it won't blow up your VRAM the way a long prompt does on an older model. That combination of a small footprint, a generous context that stays cheap, and a permissive licence is why a 9B like this one is often the right first model rather than a stepping stone to something bigger.
Which GPU runs it
This is an entry-friendly model: a 12GB RTX 3060 12GB is the value pick and even 8GB cards run the Q4 build. See best GPU for Qwen3.5 9B for the live-price ranking.
Referral link
No spare GPU? Rent a 12GB one
A 12GB RTX 3060 was listing from about $0.036/hr on Vast.ai's spot market (26 Sep 2026) — plenty for Qwen3.5 9B at Q8. Prices move; check the live figure first.
Referral link — we may earn a commission at no cost to you. Prices on our compare page are the provider's live figures, cheapest first; this never changes the ranking.
How to run it
The whole command is the default tag:
ollama run qwen3.5 # pulls the 9B
# or be explicit:
ollama run qwen3.5:9b
Both pull the same 9B. That default is exactly why it's worth knowing what you're getting — if you wanted a different size, you'd have to ask for it (qwen3.5:4b, qwen3.5:27b, and so on). For llama.cpp, use a GGUF of Qwen3.5-9B with llama-server.
Who should skip it
Skip Qwen3.5 9B if you have a 24GB card idle — you can run a 30B-class model for noticeably more capability at that size. Skip it if you specifically need vision or audio; Qwen3.5 9B is a text model, whereas Gemma 4 12B is multimodal for a similar footprint. And if you're on a tiny 6GB card, drop to a smaller Qwen (the 4B) rather than squeezing this one.
Qwen3.5 9B: common questions
What model does 'ollama run qwen3.5' download?
The 9B. On ollama.com/library, 9b is the default tag the bare qwen3.5 name resolves to.
How much VRAM does Qwen3.5 9B need?
About 6.7 GB at Q4_K_M and 10.8 GB at Q8_0 (computed by our VRAM engine at 16K context). Q4 fits an 8GB card.
What's the best GPU for Qwen3.5 9B?
A 12GB card like the RTX 3060 12GB is the value pick; 8GB cards run the Q4 build. See our best-GPU page for the live ranking.
Is Qwen3.5 9B free for commercial use?
Yes — the model card lists it under the Apache-2.0 licence.
If you're weighing sizes across the Qwen line, our best GPU for local AI guide maps VRAM to model size — and if you later outgrow the 9B, moving up to a 24GB card opens the door to the Qwen3.6 35B-A3B mixture-of-experts model, which gives you far more capability while staying quick to run.
