GPUs · NVIDIA
Can the GTX 1650 run local LLMs?
With 4 GB of VRAM (about 3.6 GB usable) and 128 GB/s of memory bandwidth, the GTX 1650 runs 5 of 39 popular open models fully on the GPU at 8K context. The largest is Phi-4 Mini 3.8B at Q4_K_M, at roughly 37 tokens per second.
Best picks for the GTX 1650
Every model on the GTX 1650
Results at 8K context, picking the highest quality quantization that fits. "Offload" means part of the model runs from system RAM (we assume 32 GB), which works but is much slower.
| Model | Status | Quant | Memory | Speed |
|---|---|---|---|---|
| Phi-4 Mini 3.8B | Runs well | Q4_K_M | 3.2 GB | ~37 tok/s |
| Llama 3.2 3B | Runs well | Q5_K_M | 3.4 GB | ~37 tok/s |
| Llama 3.2 1B | Runs well | Q8_0 | 1.6 GB | ~65 tok/s |
| Gemma 3 1B | Runs well | Q8_0 | 1.4 GB | ~78 tok/s |
| Qwen3 0.6B | Runs well | Q8_0 | 0.9 GB | ~130 tok/s |
| Qwen3 4B | Offload | Q4_K_M | 3.7 GB | ~5.6 tok/s |
| Qwen3.5 4B | Offload | Q4_K_M | 3.9 GB | ~5.2 tok/s |
| Gemma 3 4B | Offload | Q4_K_M | 4.7 GB | ~7.0 tok/s |
| Mistral 7B v0.3 | Offload | Q4_K_M | 5.7 GB | ~4.9 tok/s |
| Qwen2.5-Coder 7B | Offload | Q4_K_M | 5.8 GB | ~4.7 tok/s |
| Llama 3.1 8B | Offload | Q4_K_M | 6.2 GB | ~4.7 tok/s |
| DeepSeek-R1 Distill 8B | Offload | Q4_K_M | 6.2 GB | ~4.7 tok/s |
| Qwen3 8B | Offload | Q4_K_M | 6.4 GB | ~4.6 tok/s |
| Qwen3.5 9B | Offload | Q4_K_M | 7.0 GB | ~4.2 tok/s |
| Mistral Nemo 12B | Offload | Q4_K_M | 9.2 GB | ~3.6 tok/s |
| Gemma 4 12B | Offload | Q4_K_M | 11 GB | ~3.9 tok/s |
| Phi-4 14B | Offload | Q4_K_M | 11 GB | ~3.2 tok/s |
| Gemma 3 12B | Offload | Q4_K_M | 11 GB | ~3.8 tok/s |
| Qwen3 14B | Offload | Q4_K_M | 11 GB | ~3.2 tok/s |
| DeepSeek-R1 Distill 14B | Offload | Q4_K_M | 11 GB | ~3.2 tok/s |
| gpt-oss 20B | Offload | Q4_K_M | 14 GB | ~11 tok/s |
| Codestral 22B | Offload | Q4_K_M | 16 GB | ~2.3 tok/s |
| Devstral Small 24B | Offload | Q4_K_M | 17 GB | ~2.1 tok/s |
| Mistral Small 3 24B | Offload | Q4_K_M | 18 GB | ~2.1 tok/s |
| Gemma 4 26B-A4B (MoE) | Offload | Q4_K_M | 19 GB | ~10.0 tok/s |
| Qwen3.6 27B | Offload | Q4_K_M | 19 GB | ~1.9 tok/s |
| Qwen3.6 35B-A3B (MoE) | Offload | Q3_K_M | 19 GB | ~14 tok/s |
| Qwen3 32B | Offload | Q3_K_M | 19 GB | ~2.0 tok/s |
| DeepSeek-R1 Distill 32B | Offload | Q3_K_M | 19 GB | ~2.0 tok/s |
| Qwen2.5-Coder 32B | Offload | Q3_K_M | 19 GB | ~2.0 tok/s |
| Qwen3 30B-A3B (MoE) | Offload | Q4_K_M | 21 GB | ~11 tok/s |
| Qwen3-Coder 30B-A3B | Offload | Q4_K_M | 21 GB | ~11 tok/s |
| Gemma 4 31B | Offload | Q3_K_M | 21 GB | ~2.1 tok/s |
| Gemma 3 27B | Offload | Q4_K_M | 22 GB | ~1.9 tok/s |
| Llama 3.3 70B | Won’t fit | — | 48 GB | — |
| DeepSeek-R1 Distill 70B | Won’t fit | — | 48 GB | — |
| Llama 4 Scout 109B-A17B | Won’t fit | — | 73 GB | — |
| gpt-oss 120B | Won’t fit | — | 77 GB | — |
| Qwen3 235B-A22B (MoE) | Won’t fit | — | 155 GB | — |
What this means in practice
The GTX 1650 is comfortable with models up to about 4B parameters. gpt-oss 20B is the most useful step up beyond that: it needs about 14 GB, so part of it runs from system RAM, but as a mixture-of-experts model it stays reasonably quick at around 11 tokens per second. Models from Llama 3.3 70B upward need 48 GB or more and are out of reach. NVIDIA cards have the broadest support: Ollama, LM Studio and llama.cpp all use CUDA out of the box.
The next memory tier up is 6 GB: GTX 1660 / 1660 Super.
Compare the GTX 1650
Questions
What is the biggest model the GTX 1650 can run?
At 8K context the largest model that fits entirely in its 4 GB of VRAM is Phi-4 Mini 3.8B at Q4_K_M, at roughly 37 tokens per second.
Can the GTX 1650 run 70B models like Llama 3.3 70B?
No. It needs about 48 GB at Q4, which is far beyond 4 GB even with system RAM offload.
What is the best model for coding on the GTX 1650?
Coding models that fit at full speed are limited at 4 GB. Small general models still help with short snippets.
Estimates based on memory size and bandwidth; real speed varies with drivers, software and context length. See how we calculate VRAM.