LLM on Home Gaming PC | Exact Models by GPU (2026)

10-Minute Quick Start

Install Ollama (Windows / macOS / Linux): ollama.com/download
Pick a model from the table below based on your VRAM.
Run it in a terminal, e.g. ollama run qwen3:14b
Want a GUI? Install LM Studio or Open WebUI.

Pick by GPU (2026)

Numbers assume Q4_K_M quantization via Ollama / llama.cpp, 8k context, Windows 11 or Linux with recent NVIDIA drivers. Tokens/sec are real-world decode speeds, not theoretical peaks.

GPU	VRAM	Best pick	Command	Speed
RTX 3050 / 4050 laptop	6 GB	Qwen3 4B	`ollama run qwen3:4b`	~55 tok/s
RTX 3060 / 4060 (8 GB)	8 GB	Qwen3 8B	`ollama run qwen3:8b`	~45 tok/s
RTX 3060 12 GB	12 GB	Qwen3 14B (Q4)	`ollama run qwen3:14b`	~28 tok/s
RTX 4070 / 5070	12 GB	Qwen3 14B (Q4)	`ollama run qwen3:14b`	~55 tok/s
RTX 4060 Ti / 4070 Ti Super	16 GB	Qwen3 14B (Q5) or Gemma 3 27B (Q3)	`ollama run qwen3:14b-instruct-q5_K_M`	~48 tok/s
RTX 4080 / 5080	16 GB	Qwen3 14B (Q5), DeepSeek R1 Distill 14B	`ollama run deepseek-r1:14b`	~70 tok/s
RTX 3090 / 4090	24 GB	Qwen3.8 27B (Q4) or Gemma 4 31B (Q3)	`ollama run qwen3.8:27b`	~35 tok/s (27B)
RTX 5090	32 GB	Qwen3.8 27B (Q5) or Gemma 4 31B (Q4)	`ollama run qwen3.8:27b`	~22 tok/s (27B Q5)
Dual RTX 3090 / 4090	48 GB	Qwen3.8 27B (Q8) or Gemma 4 31B (Q8)	`ollama run qwen3.8:27b`	~18 tok/s

What to Run for Each Job

Coding assistant

12 GB: ollama run qwen2.5-coder:14b

16 GB: ollama run qwen2.5-coder:14b-instruct-q5_K_M

24 GB: ollama run qwen2.5-coder:32b

Qwen2.5-Coder is the strongest open coder in its size class. Pair with Continue.dev in VS Code for a Copilot-style setup.

Reasoning & math

8 GB: ollama run deepseek-r1:8b

16 GB: ollama run deepseek-r1:14b

24 GB: ollama run deepseek-r1:32b

DeepSeek R1 distills show visible chain-of-thought. Great for step-by-step problems; slower than plain instruct models.

General chat & writing

8 GB: ollama run qwen3:8b

12 GB: ollama run gemma4:12b

24 GB: ollama run gemma4:31b

Gemma 4 is Apache-2.0-friendly and neutral in tone. Qwen3 8B is the strongest all-rounder at 8 GB.

Vision (image + text)

12 GB: ollama run gemma4:12b

24 GB: ollama run qwen3.8:27b

Good for OCR, screenshot Q&A, and describing local photos entirely offline.

Nordic / multilingual

12 GB: ollama run qwen3:8b

24 GB: ollama run gemma4:31b

Qwen3.8 and Gemma 4 handle Norwegian, Swedish, and Danish acceptably. Verify on your own prompts — small models still stumble on idioms.

Uncensored / roleplay

12 GB: ollama run dolphin-mixtral:8x7b

24 GB: ollama run nous-hermes2:34b

Community fine-tunes with fewer refusals. Use responsibly — you still own the output.

VRAM Rule of Thumb

To estimate VRAM needed for a model at Q4: parameters × 0.6 GB. Add ~1-2 GB for context (8k) and overhead.

7B Q4 → ~5 GB VRAM
14B Q4 → ~9 GB VRAM
32B Q4 → ~20 GB VRAM
70B Q4 → ~42 GB VRAM (needs dual GPU or Q2/Q3)

If it does not fit, drop to Q3 (small quality loss) or Q2 (noticeable quality loss) before dropping a model tier.

Buying a New Rig? Two Recipes.

Budget builder (~$1,500)

GPU: RTX 5070 12 GB
CPU: Ryzen 7 7700
RAM: 32 GB DDR5-6000
SSD: 1 TB NVMe Gen4

Runs Qwen3 14B at ~55 tok/s. Plays every 2026 game at 1440p high.

Serious local LLM (~$3,500)

GPU: RTX 5090 32 GB
CPU: Ryzen 9 7950X3D
RAM: 64 GB DDR5-6000
SSD: 2 TB NVMe Gen4

Runs Qwen3.8 27B comfortably at Q5. Room for 256k context on smaller models.

Common Traps

Do not use bf16/fp16 weights on consumer GPUs — you will waste VRAM. Stick to GGUF Q4/Q5.
Context length costs VRAM. 32k context can double the memory of a small model. Start at 8k.
Windows shares VRAM with the desktop. Close Chrome and games before loading big models, or lose 1-2 GB usable VRAM.
Laptop RTX ≠ desktop RTX. An RTX 4070 laptop has 8 GB; the desktop card has 12 GB. Check your actual VRAM with nvidia-smi.
CPU-only is possible but slow. Qwen3 4B on CPU gets ~5-8 tok/s. Fine for background jobs, painful for chat.

Quick Buying Rule

Gaming + casual LLM: aim for 12 GB VRAM (RTX 5070 or 4070 Super). Serious local coding/reasoning: aim for 24-32 GB VRAM (RTX 4090 or 5090).