Only with CPU offload

Qwen3.5 9B at Q4_K_M needs 6.3 GB but only 5.3 GB is addressable, so about 18% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 21 tokens per second — usable for batch work, painful for chat.

82% on GPU 8K context Q4_K_M · 5.4 GB Apache 2.0 Released 28 Feb 2026 Vision

The default for 8–12 GB cards in 2026: beats every older 8B on every published benchmark, with vision.

Hardware requirements What hardware do I need to run Qwen3.5 9B? →

At Q4_K_M with an 8K context: a 10 GB GPU (Arc B570) or a 16 GB Mac (M4). Minimum, recommended, high-quality and long-context tiers, with all 114 devices checked at every quantisation.

Fits instead: Qwen3.5 4B (3.5 GB) · Qwen3.5 2B (2.0 GB)

The VRAM budget

weights 5.4 GB
Weights 5.4 GB KV cache @ 8K 0.25 GB Runtime overhead 0.6 GB Over budget 1.0 GB past 5.3 GB

Quantisation ladder

QuantWeightsTotal @ 8KMax contextTok/sQualityFit
F16 18.0 GB 18.8 GB — ~2.5 Reference 13.5 GB over
Q8_0 9.5 GB 10.4 GB — ~6.2 −0.1% ppl 5.1 GB over
Q6_K 7.4 GB 8.2 GB — ~9.8 −0.4% ppl 2.9 GB over
Q5_K_M 6.4 GB 7.2 GB — ~13 −0.8% ppl 1.9 GB over
Q4_K_M 5.4 GB 6.3 GB — ~21 −1.9% ppl 1.0 GB over
Q3_K_M 4.4 GB 5.2 GB 9K 46 −5.4% ppl Fits
Q2_K 3.8 GB 4.6 GB 29K 54 −15% ppl Long context

Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. Only 8 of its 32 blocks keep a per-token KV cache; the rest are linear-attention, Mamba or convolution blocks with a fixed-size state.

How to run it

terminal
$ llama-server \
    -hf Qwen/Qwen3.5-9B:Q4_K_M \
    -c 8192 -ngl 26

The engine underneath most of the others. Every knob is exposed. More on llama.cpp.

01Download is 5.4 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.
03If it falls back to CPU silently, drop the context first, then step down a quantisation.
See all models for this rig Compare with another model

Related models on a GeForce GTX 1660 Super

ModelTotalTok/s est.
Qwen3.5 4B
4.66B · same family
Q4_K_M · 3.5 GB 78 Runs great
Qwen3.5 2B
2.27B · same family
Q4_K_M · 2.0 GB 159 Runs great
Qwen3.5 0.8B
0.87B · same family
Q4_K_M · 1.2 GB 416 Runs great
Qwen3.5 122B-A10B
125B MoE · same family
Q4_K_M · 71.1 GB — Won't fit
Ornith 1.5 9B
9.41B · similar size
Q4_K_M · 6.1 GB ~22 CPU offload
Ministral 3 8B
8.92B · similar size
Q4_K_M · 6.7 GB ~18 CPU offload
Granite 4.1 8B
8.79B · similar size
Q4_K_M · 6.8 GB ~17 CPU offload

Every model against the GeForce GTX 1660 Super →

Qwen3.5 9B on other hardware

DeviceTotalTok/s est.
GeForce RTX 4060 Laptop
7.0 GB usable · 256 GB/s
6.3 GB 29 Tight fit
Arc B570
8.8 GB usable · 380 GB/s
6.3 GB 42 Runs great
GeForce RTX 3060 12 GB
10.6 GB usable · 360 GB/s
6.3 GB 40 Runs great
M4 · 16 GB
10.7 GB usable · 120 GB/s
6.3 GB 12 Runs great
GeForce RTX 4090
22.4 GB usable · 1008 GB/s
6.3 GB 112 Runs great
Radeon RX 7900 XTX
22.4 GB usable · 960 GB/s
6.3 GB 107 Runs great
M3 Max · 64 GB
48.0 GB usable · 400 GB/s
6.3 GB 41 Runs great

All 114 devices, by tier →