can it run
4.6 GB vs 4 GB usable — partial CPU offload, expect large speed loss With a q8 (quantized) KV cache it comes to the edge (4.3GB) — possible with modest context, at some quality cost.
4GB · Turing · model file 4.1GB · 7B
computed roofline: 192 GB/s × 0.48–0.72 efficiency window / 4.1GB (Q4_K_M) — bands, never points.
→ Tesla M60 (16GB, used) — cheapest catalog machine that runs mistral 7B fully ($35)
→ gtx-1660-super — the next step up from NVIDIA GeForce GTX 1650 Super
→ every model the NVIDIA GeForce GTX 1650 Super can run — full list
→explore the full catalog91 machines indexed · live prices · what each one can run