Question 1

What models can the NVIDIA GeForce GTX 1660 Ti 6GB run?

Accepted Answer

With 6GB of VRAM: qwen3:0.6b, qwen3:1.7b, qwen3:4b, qwen3:8b, qwen2.5vl:7b, llama3.1:8b, llama3.2:1b, llama3.2:3b run fully in VRAM. No tensor cores despite Turing arch — llama.cpp CUDA falls back to dp4a INT8 kernels; build with GGML_CUDA_FORCE_MMQ=ON (CMAKE_CUDA_ARCHITECTURES=75) or you lose ~20% speed 6GB VRAM is tight: dense 7B Q4_K_M (~4.1GB weights) fits with only ~8K context; realistic lane is small dense models or MoE with --n-cpu-moe offload Bench evidence is MoE-with-CPU-offload (28 tok/s depends heavily on host CPU/RAM bandwidth for expert matmuls); no clean sourced dense-7B CUDA median found — dense 7B fully on-GPU will be slower than the cited MoE figure per-token on prompt eval (~43 tok/s prefill measured) Laptop variants (Max-Q) downclock significantly; laptop PCIe may downshift to x8 No NVLink; multi-GPU scaling is limited to llama.cpp tensor split over PCIe

Question 2

How fast is the NVIDIA GeForce GTX 1660 Ti 6GB for local LLMs?

Accepted Answer

computed band 27.0-40.0 tok/s for 8B Q4_K_M (roofline, 288.0 GB/s VRAM); community: 27-28 tok/s Qwen3.6 35B-A3B (MoE, ~3B active, --n-cpu-moe offload) Q4_K_M (https://pub.towardsai.net/a-gpu-poors-guide-to-local-llm-inference-in-2026-48d59cafd215)

Question 3

Is the NVIDIA GeForce GTX 1660 Ti 6GB still worth using for local LLMs in 2026?

Accepted Answer

computed band 27.0-40.0 tok/s for 8B Q4_K_M (roofline, 288.0 GB/s VRAM); community: 27-28 tok/s Qwen3.6 35B-A3B (MoE, ~3B active, --n-cpu-moe offload) Q4_K_M (https://pub.towardsai.net/a-gpu-poors-guide-to-local-llm-inference-in-2026-48d59cafd215) No tensor cores despite Turing arch — llama.cpp CUDA falls back to dp4a INT8 kernels; build with GGML_CUDA_FORCE_MMQ=ON (CMAKE_CUDA_ARCHITECTURES=75) or you lose ~20% speed 6GB VRAM is tight: dense 7B Q4_K_M (~4.1GB weights) fits with only ~8K context; realistic lane is small dense models or MoE with --n-cpu-moe offload Bench evidence is MoE-with-CPU-offload (28 tok/s depends heavily on host CPU/RAM bandwidth for expert matmuls); no clean sourced dense-7B CUDA median found — dense 7B fully on-GPU will be slower than the cited MoE figure per-token on prompt eval (~43 tok/s prefill measured) Laptop variants (Max-Q) downclock significantly; laptop PCIe may downshift to x8 No NVLink; multi-GPU scaling is limited to llama.cpp tensor split over PCIe For serious work, the upgrade path is: rtx-3060-ti, rtx-3070.

Best local LLMs for the NVIDIA GeForce GTX 1660 Ti 6GB

Speed

What fits (computed)

Cited community benches (1 rows)

Runs fully in memory (machine alternatives)

Borderline — runs, but offloads

Gotchas

Where to go next