Yes, just — 9.2 GB spare
GLM-5.3-Flash 320B-A18B at Q4_K_M fits your Instinct MI300X entirely on the GPU at 8K context, at an estimated 122 tokens per second. Past 866K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
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.
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 317.6 GB | 318.3 GB | — | ~1.9 | −0.1% ppl | 127.9 GB over |
| Q6_K | 245.1 GB | 245.8 GB | — | ~4.3 | −0.4% ppl | 55.4 GB over |
| Q5_K_M | 211.9 GB | 212.6 GB | — | ~10 | −0.8% ppl | 22.2 GB over |
| Q4_K_M | 180.5 GB | 181.2 GB | 866K | 122 | −1.9% ppl | Recommended |
| Q3_K_M | 146.1 GB | 146.8 GB | 1024K | 151 | −5.4% ppl | Long context |
| Q2_K | 125.2 GB | 125.9 GB | 1024K | 176 | −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 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
$ llama-server \
-hf zai-org/GLM-5.3-Flash:Q4_K_M \
-c 8192 -ngl 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.