Yes, just — 7.3 GB spare

Qwen3.5 122B-A10B at Q4_K_M fits your H100 SXM entirely on the GPU at 8K context, at an estimated 139 tokens per second. There is room for its full 256K window.

Fully on GPU 8K context Q4_K_M · 70.3 GB Apache 2.0 Released 24 Feb 2026 Vision

The 96–128 GB unified-memory model: 122B of knowledge at 10B-active speed.

What hardware do I need for Qwen3.5 122B-A10B? →

The VRAM budget

weights 70.3 GB

Weights 70.3 GB KV cache @ 8K 0.19 GB Runtime overhead 0.6 GB Free 7.3 GB of 78.4 GB

Quantisation ladder

Quant	Weights	Total @ 8K	Max context	Tok/s	Quality	Fit
Q8_0	123.7 GB	124.5 GB	—	~3.7	−0.1% ppl	46.1 GB over
Q6_K	95.5 GB	96.2 GB	—	~9.1	−0.4% ppl	17.8 GB over
Q5_K_M	82.5 GB	83.3 GB	—	~28	−0.8% ppl	4.9 GB over
Q4_K_M	70.3 GB	71.1 GB	256K	139	−1.9% ppl	Recommended
Q3_K_M	56.9 GB	57.7 GB	256K	171	−5.4% ppl	Long context
Q2_K	48.7 GB	49.5 GB	256K	200	−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 12 of its 48 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

Ollama llama.cpp LM Studio vLLM

terminal

$ llama-server \
    -hf Qwen/Qwen3.5-122B-A10B:Q4_K_M \
    -c 8192 -ngl 99

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

01Download is 70.3 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 7.3 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 256K context.

See all models for this rig Compare with another model