Yes — with 29.7 GB to spare
DeepSeek V4 Flash 284B-A13B at Q4_K_M fits your Instinct MI300X entirely on the GPU at 8K context, at an estimated 169 tokens per second. Past 637K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
The V4 that 128 GB machines can actually run at Q3. Cache is modelled as a 576-wide latent; V4 compresses it further at long context, so this is conservative.
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 281.0 GB | 282.0 GB | — | ~3.3 | −0.1% ppl | 91.6 GB over |
| Q6_K | 216.9 GB | 217.9 GB | — | ~10 | −0.4% ppl | 27.5 GB over |
| Q5_K_M | 187.5 GB | 188.4 GB | 49K | 144 | −0.8% ppl | Long context |
| Q4_K_M | 159.7 GB | 160.7 GB | 637K | 169 | −1.9% ppl | Recommended |
| Q3_K_M | 129.3 GB | 130.3 GB | 1024K | 209 | −5.4% ppl | Long context |
| Q2_K | 110.8 GB | 111.7 GB | 1024K | 243 | −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. 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 deepseek-ai/DeepSeek-V4-Flash-0731:Q4_K_M \
-c 8192 -ngl 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.