Yes, just — 5.2 GB spare
Qwen3 235B-A22B at Q4_K_M fits your H200 SXM entirely on the GPU at 8K context, at an estimated 90 tokens per second. Past 36K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Fully on GPU
8K context
Q4_K_M · 132.1 GB
Apache 2.0
Released Apr 2025
Workstation class. Realistically a 192 GB unified-memory or multi-GPU model.
The VRAM budget
weights 132.1 GB
Weights 132.1 GB
KV cache @ 8K 1.47 GB
Runtime overhead 0.6 GB
Free 5.2 GB of 139.4 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 232.5 GB | 234.6 GB | — | ~1.5 | −0.1% ppl | 95.2 GB over |
| Q6_K | 179.5 GB | 181.5 GB | — | ~3.4 | −0.4% ppl | 42.1 GB over |
| Q5_K_M | 155.1 GB | 157.2 GB | — | ~7.7 | −0.8% ppl | 17.8 GB over |
| Q4_K_M | 132.1 GB | 134.2 GB | 36K | 90 | −1.9% ppl | Recommended |
| Q3_K_M | 107.0 GB | 109.0 GB | 128K | 112 | −5.4% ppl | Long context |
| Q2_K | 91.6 GB | 93.7 GB | 128K | 130 | −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.
How to run it
$ llama-server \
-hf Qwen/Qwen3-235B-A22B:Q4_K_M \
-c 8192 -ngl 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 132.1 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.
03Only 5.2 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 36K context.