Not sure which model to run? This guide cuts through the noise. We recommend the single best model per VRAM tier — the one most users should start with.
4 GB VRAM (Budget GPUs / iGPUs)
Best pick: Phi-4 Mini (3.8B, Q4_K_M)
- VRAM needed: 3.2 GB
- Speed: ~35–50 t/s on RTX 4060, ~20 t/s on integrated graphics
- Ollama:
ollama run phi4-mini - Why: Microsoft's Phi-4 Mini outperforms models twice its size on reasoning and coding tasks. Fits in 3.2 GB, leaving headroom for context.
Runner-up: Gemma 3 4B (ollama run gemma3:4b) — better for creative writing, 3.5 GB.
What to expect: Capable assistant for everyday chat, coding help, and writing. Don't expect 70B-level reasoning, but for 4 GB, this is remarkable.
8 GB VRAM (RTX 4060, RX 7600, laptop GPUs)
Best pick: Llama 3.1 8B (Q4_K_M)
- VRAM needed: 5.8 GB
- Speed: 55–90 t/s on RTX 4060 / 4070
- Ollama:
ollama run llama3.1:8b - Why: Meta's Llama 3.1 8B is the most widely tested 8B model. Strong at all tasks, extensive community support, fast on 8 GB cards.
Runner-up for coding: Qwen 2.5 Coder 7B (ollama run qwen2.5-coder:7b) — best pure-coding 8B model.
Runner-up for reasoning: DeepSeek R1 8B Distill (ollama run deepseek-r1:8b) — near-GPT-4-level reasoning in 8B.
12 GB VRAM (RTX 4070, RTX 3060 12GB, RX 7700 XT)
Best pick: Qwen 3 8B (Q8_0) — run at higher quality since you have headroom
- VRAM needed: 9.5 GB
- Speed: 60–95 t/s
- Ollama:
ollama run qwen3:8b - Why: With 12 GB VRAM, you can run 8B models at Q8_0 (near-lossless) quality instead of Q4, meaningfully improving outputs.
Or try: Gemma 3 12B Q4 (ollama run gemma3:12b) — Google's best mid-size model.
16 GB VRAM (RTX 4060 Ti 16GB, RTX 4070 Ti, RX 7800 XT)
Best pick: Qwen 3 14B (Q4_K_M)
- VRAM needed: 9.8 GB (leaves 6 GB for context)
- Speed: 64–98 t/s
- Ollama:
ollama run qwen3:14b - Why: 14B models hit a quality sweet spot. Qwen 3 14B is state-of-the-art at this size — beats GPT-3.5 on most tasks.
Runner-up for coding: Qwen 2.5 Coder 14B (ollama run qwen2.5-coder:14b)
Best for long documents: Mistral Small 3.1 (ollama run mistral-small) — 128K context window.
24 GB VRAM (RTX 4090, RTX 3090, RX 7900 XTX)
Best pick: DeepSeek R1 32B (Q4_K_M)
- VRAM needed: 19.5 GB
- Speed: 45–55 t/s
- Ollama:
ollama run deepseek-r1:32b - Why: DeepSeek R1 32B is a reasoning powerhouse that fits in 24 GB and rivals GPT-4o on math, coding, and logic. A game-changer for 24 GB GPU owners.
Alternative: Qwen 3 32B (ollama run qwen3:32b) — also fits, stronger at general tasks.
32 GB+ VRAM (RTX 5090, Apple M4 Max, Professional GPUs)
Best pick: Llama 4 Scout (Q4_K_M)
- VRAM needed: 12–14 GB (leaves 18+ GB for a massive context window)
- Speed: 175 t/s on RTX 5090
- Ollama:
ollama run llama4:scout - Why: With 32 GB VRAM, you can run Llama 4 Scout (17B active parameters, 10M context) at blazing speed with enormous context — useful for long documents, codebases, or multi-turn conversations.
For 70B models: Llama 3.3 70B Q4_K_M needs 42 GB — works on Apple M2/M4 Ultra (128–192 GB unified memory) or dual RTX 4090/5090 setups.
The Quick Reference Table
| VRAM | Best Model | Ollama Command | Speed |
|---|---|---|---|
| 4 GB | Phi-4 Mini | ollama run phi4-mini | ~40 t/s |
| 8 GB | Llama 3.1 8B | ollama run llama3.1:8b | ~70 t/s |
| 12 GB | Qwen 3 8B Q8 | ollama run qwen3:8b | ~80 t/s |
| 16 GB | Qwen 3 14B | ollama run qwen3:14b | ~85 t/s |
| 24 GB | DeepSeek R1 32B | ollama run deepseek-r1:32b | ~50 t/s |
| 32 GB+ | Llama 4 Scout | ollama run llama4:scout | 175 t/s |
To size a specific model, use the VRAM calculator: the memory it needs at every quantization and context length.