Best Ollama Models for 16GB VRAM (16GB VRAM)
gpt-oss:20b and the 14B class at Q5 — what 16GB unlocks
The best Ollama model for 16GB VRAM in 2026 is gpt-oss:20b — OpenAI’s open-weight MoE at ~14–16GB, generating at small-model speed because only ~3.6B params are active per token. Qwen 3 14B at Q5 (~11GB) is the dense alternative with context headroom, and Devstral 24B (~15GB) the agentic-coding pick. 32B models still need 24GB.
Models that fit in 16GB VRAM
Tested reference: RTX 4080 / RTX 4060 Ti 16GB. Figures are for Q4_K_M with a modest context window.
| Model | Size | Build | VRAM used | Speed |
|---|---|---|---|---|
| WINNERgpt-oss:20b OpenAI’s open-weight model, sized for 16GB cards — MoE, so it generates at small-model speed. A tight fit: run it alone. ollama pull gpt-oss:20b | 20.9B (MoE, ~3.6B active) | MXFP4 | ~14–16GB | ~40–50 tok/s |
| Qwen 3 14B (Q5) 16GB buys the quant headroom 12GB lacks — the 14B at Q5 with a long context is the dense sweet spot. ollama pull qwen3:14b-q5_K_M | 14B | Q5_K_M | ~11GB | ~30 tok/s |
| Devstral 24B Purpose-built agentic coder (multi-file edits) — fits 16GB only with a short context. ollama pull devstral:24b | 24B | Q4_K_M | ~15GB | ~18–22 tok/s |
| DeepSeek-R1 14B Chain-of-thought reasoning with plenty of KV headroom at this tier. ollama pull deepseek-r1:14b | 14B | Q4_K_M | ~9.0GB | ~33 tok/s |
| Phi-4 14B STEM/reasoning specialist — pairs well as the “second model” beside a loaded coder. ollama pull phi4 | 14B | Q4_K_M | ~9.0GB | ~34 tok/s |
What won't fit in 16GB
- ✗Qwen3.6 27B (needs ~17GB) — The current 24GB-tier headline pick — a hair over 16GB at Q4.
- ✗Qwen 2.5 Coder 32B (needs ~20GB) — The SOTA local coder needs a 24GB card.
How to fit more in 16GB
- →gpt-oss:20b is a tight fit — load it with nothing else on the GPU, keep num_ctx modest, and confirm `ollama ps` reports 100% GPU.
- →Use 16GB’s headroom for quant, not just size: a 14B at Q5_K_M (~11GB) plus a long context usually beats a 22–24B dense model squeezed in at Q4.
- →Dense 22–24B models (Devstral, Codestral) fit only with short contexts — the MoE gpt-oss:20b gives you 20B-class reasoning at roughly twice their tok/s.
Quick start
curl -fsSL https://ollama.com/install.sh | shollama run gpt-oss:20bGo from “it runs” to actually building
Learn to run local models, wire up RAG, and build agents — hands-on, with a free first chapter of every course.
Start learning free →Frequently asked questions
What are the best GGUF models for 16GB VRAM?
gpt-oss:20b (~14–16GB, ships in MXFP4) is the flagship; Qwen 3 14B at Q5_K_M (~11GB) and Devstral 24B at Q4 (~15GB) round out the tier. Ollama pulls are GGUF under the hood, so the same picks apply if you load them in LM Studio.
Can 16GB VRAM run gpt-oss:20b?
Yes — that is the card it was sized for. OpenAI ships it in MXFP4 at ~14–16GB, and because it is MoE (~3.6B active params per token) it generates far faster than a dense 20B. Keep desktop GPU use light while it is loaded.
Should I squeeze a 32B model onto 16GB?
Usually no — Qwen 2.5 32B only fits at Q3 (~15GB) with a short context at ~14 tok/s, and a 14B at Q4/Q5 is faster and typically better on real tasks. For full-quality 32B, this tier’s ceiling is real: you want 24GB.
Related guides
Ready to Go Beyond Tutorials?
20 structured courses with hands-on chapters - build RAG chatbots, AI agents, and ML pipelines on your own hardware.
Was this helpful?