Can I run Apertus-70B-Instruct-2509?

Apertus-70B-Instruct-2509 needs about 50.2 GB of memory at Q4_K_M, and the cheapest GPU we track that holds it is the NVIDIA RTX Pro 6000 Blackwell 96GB (96 GB, $8565 list price). 7 of the 82 devices we track run it fully at Q4. These are estimates, not measured benchmarks.

70.6B parameters, by swiss-ai. License: apache-2.0. Model card (params from HF API, retrieved 2026-11-04) · How much VRAM for 70B models · Q4 vs Q8 vs FP16

Q4_K_M needs

50.2 GB 42.8 GB weights

Q8_0 needs

87.3 GB 75.0 GB weights

FP16 needs

163.4 GB 141.2 GB weights

7 of 82 tracked GPUs and Macs run it fully at Q4.

Memory by context length

Longer chats and documents need more memory for the KV cache. All 80 layers keep a full cache. Sizes assume an FP16 KV cache; a Q8 cache roughly halves the cache part.

From the model's config.json (80 full-attention layers × 8 KV heads × 128 head size, retrieved 2026-11-04). Check the model card for its maximum supported context. Try other sizes in the VRAM calculator or compare models in our context length guide.

Which hardware runs it

We may earn a commission when you buy through these links, at no extra cost to you.