With 16 GB VRAM, assuming 16 GB system RAM for partial offload, these are the best open models that fit. The top pick is gpt-oss 20B (MoE, MXFP4) at MXFP4 (~15.1 GB, ~89 tok/s): ollama run gpt-oss:20b.
Best for coding: gpt-oss 20B (MoE, MXFP4) (gpt-oss:20b), DeepSeek Coder V2 Lite 16B (MoE) (deepseek-coder-v2:16b-lite-instruct-q5_K_M), Qwen3-Coder 30B-A3B (MoE) (qwen3-coder:30b).
Runtimes available for this hardware: Ollama, LM Studio, llama.cpp, vLLM.
Open the interactive checker for RTX 4080 16GB| # | Model | Quant | Memory needed | Fit | Speed | Command |
|---|---|---|---|---|---|---|
| 1 | gpt-oss 20B (MoE, MXFP4) | MXFP4 | 15.1 GB | fits in fast memory | ~89 tok/s | ollama run gpt-oss:20b |
| 2 | DeepSeek R1 Distill Qwen 14B | Q4_K_M | 10.3 GB | fits in fast memory | ~48 tok/s | ollama run deepseek-r1:14b |
| 3 | Phi-4 14B | Q4_K_M | 10.4 GB | fits in fast memory | ~47 tok/s | ollama run phi4:14b |
| 4 | Qwen3 14B | Q4_K_M | 10.6 GB | fits in fast memory | ~46 tok/s | ollama run qwen3:14b |
| 5 | Mistral Nemo 12B | Q8_0 | 14.2 GB | fits in fast memory | ~33 tok/s | ollama run mistral-nemo:12b-instruct-2407-q8_0 |
| 6 | Gemma 3 12B | Q8_0 | 14.2 GB | fits in fast memory | ~33 tok/s | ollama run gemma3:12b-it-q8_0 |
| 7 | Llama 3.2 Vision 11B | Q8_0 | 13.1 GB | fits in fast memory | ~36 tok/s | ollama run llama3.2-vision:11b-instruct-q8_0 |
| 8 | Nimble v3 9B (Bespoke Labs, decision) | Q8_0 | 10.6 GB | fits in fast memory | ~45 tok/s | ollama run nimble-v3:9b-q8_0 |
| 9 | Qwen3 8B | Q8_0 | 9.9 GB | fits in fast memory | ~48 tok/s | ollama run qwen3:8b-q8_0 |
| 10 | DeepSeek R1 0528 Qwen3 8B | Q8_0 | 9.9 GB | fits in fast memory | ~48 tok/s | ollama run deepseek-r1:8b-0528-qwen3-q8_0 |
| 11 | DeepSeek R1 Distill Qwen 7B | Q8_0 | 9.1 GB | fits in fast memory | ~53 tok/s | ollama run deepseek-r1:7b-qwen-distill-q8_0 |
| 12 | Qwen 2.5 VL 7B | Q8_0 | 10.4 GB | fits in fast memory | ~46 tok/s | ollama run qwen2.5vl:7b-q8_0 |
Estimates, not benchmarks. Model sizes are ollama.com download sizes; memory needed adds headroom for ~8k tokens of context; speeds are rough figures from memory bandwidth. Data updated 2026-10-14.
A project by OnDevice Labs · llms.txt