will it local
No approved localmaxxing runs for the desktop GTX 1660 as of research date; bench rows are from a databasemart Ollama server bench and the 1660 Ti laptop (same TU116 family, +10% cores, GDDR6) — treat as class-indicative No tensor cores (GTX 16-series) — dp4a INT8 CUDA-core path 6GB VRAM: 7B Q4_K_M fits on-GPU with room for moderate context; 8B dense is tight GDDR5 at 192.1 GB/s — the 1660 Super (GDDR6, 336 GB/s) is meaningfully faster for LLM decode if the price delta is small Requires 1x 8-pin PCIe power on most AIB cards (120W TDP) CC 7.5 Turing — still supported by current CUDA and llama.cpp builds
computed band 18.0-27.0 tok/s for 8B Q4_K_M (roofline, 192.1 GB/s VRAM); community: 30-40 tok/s Llama 2 7B class (small models 1.5B-7B) Q4-class GGUF (Ollama) (https://www.databasemart.com/blog/ollama-gpu-benchmark-gtx1660); community: 27-28 tok/s Qwen3.6-35B-A3B (MoE, GTX 1660 Ti mobile sibling) Q4_K_M with --n-cpu-moe offload (https://pub.towardsai.net/a-gpu-poors-guide-to-local-llm-inference-in-2026-48d59cafd215)
Effective decode window: 0.48–0.72 of 192.1 GB/s nominal → ~92–138 GB/s effective (llama.cpp decode, Q4_K_M basis; arch window, community-calibrated)
→ 13b-q4 no 9.9 GB vs 6 GB usable — does not fit
→ 3b-q4 full 2.8 GB needed of 6 GB usable — headroom for context
→ 4b-q4 full 3.7 GB needed of 6 GB usable — headroom for context
→ 8b-q4 tight 6.1 of 6 GB usable — barely over; real with q8 KV cache (halves KV) or shorter context
→ 30-40 tok/s · Llama 2 7B class (small models 1.5B-7B) Q4-class GGUF (Ollama) · source
→ 27-28 tok/s · Qwen3.6-35B-A3B (MoE, GTX 1660 Ti mobile sibling) Q4_K_M with --n-cpu-moe offload · source
Nominal bandwidth: 192.1 GB/s — see the effective decode window above
→ qwen3:0.6b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 0.6B · 0.6GB · needs ~1GB (4k ctx)
→ qwen3:1.7b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 1.7B · 1.4GB · needs ~1.8GB (4k ctx)
→ qwen3:4b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 4B · 2.6GB · needs ~3.2GB (4k ctx)
→ llama3.2:1b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 1B · 1.3GB · needs ~1.4GB (4k ctx)
→ llama3.2:3b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 3B · 2.0GB · needs ~2.4GB (4k ctx)
→ gemma3:1b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 1B · 1.0GB · needs ~1.1GB (4k ctx)
→ phi4:mini on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) mini 3.8B · 2.5GB · needs ~3GB (4k ctx)
→ moondream:2b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 2B · 1.7GB · needs ~3.7GB (4k ctx)
→ whisper:tiny on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) tiny · 0.1GB · needs ~2.1GB (4k ctx)
→ whisper on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) base · 0.3GB · needs ~2.3GB (4k ctx)
→ whisper:small on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) small · 0.9GB · needs ~2.9GB (4k ctx)
→ whisper:medium on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) medium · 3.1GB · needs ~5.1GB (4k ctx)
→ kokoro:82m on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 82M · 0.3GB · needs ~2.3GB (4k ctx)
→ gemma3:4b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 4B · 3.3GB · needs ~3.8GB (4k ctx)
→ deepseek-r1:7b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) distill 7B · 4.7GB · needs ~4.9GB (4k ctx)
→ mistral:7b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 7B · 4.1GB · needs ~4.6GB (4k ctx)
→ qwen3:8b on Tesla P100 (16GB, used) ($135) 8B · 5.2GB · needs ~5.8GB (4k ctx)
→ qwen2.5vl:7b on Tesla P100 (16GB, used) ($135) 7B · 5.6GB · needs ~5.8GB (4k ctx)
→ llama3.1:8b on Tesla P100 (16GB, used) ($135) 8B · 4.9GB · needs ~5.4GB (4k ctx)
→ deepseek-r1:8b on Tesla P100 (16GB, used) ($135) distill 8B · 4.9GB · needs ~5.4GB (4k ctx)
→ llava:7b 7B — 6.7 GB vs 6 GB usable — partial CPU offload, expect large speed loss
⚠ No approved localmaxxing runs for the desktop GTX 1660 as of research date; bench rows are from a databasemart Ollama server bench and the 1660 Ti laptop (same TU116 family, +10% cores, GDDR6) — treat as class-indicative
⚠ No tensor cores (GTX 16-series) — dp4a INT8 CUDA-core path
⚠ 6GB VRAM: 7B Q4_K_M fits on-GPU with room for moderate context; 8B dense is tight
⚠ GDDR5 at 192.1 GB/s — the 1660 Super (GDDR6, 336 GB/s) is meaningfully faster for LLM decode if the price delta is small
⚠ Requires 1x 8-pin PCIe power on most AIB cards (120W TDP)
⚠ CC 7.5 Turing — still supported by current CUDA and llama.cpp builds