Only with CPU offload

GLM-5.3-Flash 320B-A18B at Q4_K_M needs 181.2 GB but only 7.0 GB is addressable, so about 97% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 1.4 tokens per second — usable for batch work, painful for chat.

3% on GPU 8K context Q4_K_M · 180.5 GB MIT Released 25 Aug 2026 Vision Not in the Ollama library

The first natively multimodal GLM-5 and the first hybrid: 34 linear-attention blocks and 11 sparse-attention blocks with a 512-wide latent cache, so a 1M window stays affordable. Z.ai says it beats GLM-5.2 at 18B active; MIT.

Hardware requirements What hardware do I need to run GLM-5.3-Flash 320B-A18B? →

At Q4_K_M with an 8K context: a 512 GB Mac (M3 Ultra). Minimum, recommended, high-quality and long-context tiers, with all 114 devices checked at every quantisation.

The VRAM budget

weights 180.5 GB
Weights 180.5 GB KV cache @ 8K 0.09 GB Runtime overhead 0.6 GB Over budget 174.2 GB past 7.0 GB

Quantisation ladder

QuantWeightsTotal @ 8KMax contextTok/sQualityFit
Q8_0 317.6 GB 318.3 GB — ~0.8 −0.1% ppl 311.3 GB over
Q6_K 245.1 GB 245.8 GB — ~1.0 −0.4% ppl 238.8 GB over
Q5_K_M 211.9 GB 212.6 GB — ~1.2 −0.8% ppl 205.6 GB over
Q4_K_M 180.5 GB 181.2 GB — ~1.4 −1.9% ppl 174.2 GB over
Q3_K_M 146.1 GB 146.8 GB — ~1.8 −5.4% ppl 139.8 GB over
Q2_K 125.2 GB 125.9 GB — ~2.1 −15% ppl 118.9 GB over

Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. Only 11 of its 45 blocks keep a per-token KV cache; the rest are linear-attention, Mamba or convolution blocks with a fixed-size state. This model uses multi-head latent attention, so its cache is a compressed latent rather than full K and V.

How to run it

terminal
$ llama-server \
    -hf zai-org/GLM-5.3-Flash:Q4_K_M \
    -c 8192 -ngl 1

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

01Download is 180.5 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 Radeon RX 6600 XT

ModelTotalTok/s est.
GLM-5.2 744B-A40B
753B MoE · same family
Q4_K_M · 424.7 GB — Won't fit
GLM-5.3 744B-A40B
753B MoE · same family
Q4_K_M · 424.7 GB — Won't fit
GLM-4.7-Flash 30B-A3B
31.2B MoE · same family
Q4_K_M · 18.6 GB ~11 CPU offload
DeepSeek V4 Flash 284B-A13B
284B MoE · similar size
Q4_K_M · 160.7 GB ~2.0 CPU offload
Ornith 1.5 397B-A17B
397B MoE · similar size
Q4_K_M · 224.1 GB ~1.5 CPU offload
Qwen3.5 397B-A17B
403B MoE · similar size
Q4_K_M · 227.4 GB — Won't fit
Qwen3 235B-A22B
235B MoE · similar size
Q4_K_M · 134.2 GB ~1.2 CPU offload

Every model against the Radeon RX 6600 XT →

GLM-5.3-Flash 320B-A18B on other hardware

DeviceTotalTok/s est.
GeForce RTX 4060 Laptop
7.0 GB usable · 256 GB/s
181.2 GB ~1.4 CPU offload
GeForce RTX 3060 12 GB
10.6 GB usable · 360 GB/s
181.2 GB ~1.4 CPU offload
M4 · 16 GB
10.7 GB usable · 120 GB/s
181.2 GB — Won't fit
GeForce RTX 4090
22.4 GB usable · 1008 GB/s
181.2 GB ~1.6 CPU offload
Radeon RX 7900 XTX
22.4 GB usable · 960 GB/s
181.2 GB ~1.6 CPU offload
M3 Max · 64 GB
48.0 GB usable · 400 GB/s
181.2 GB — Won't fit
M3 Ultra · 512 GB
384.0 GB usable · 800 GB/s
181.2 GB 21 Runs great

All 114 devices, by tier →