Run AI locally · from the Reel

Best open-source models you can run locally

The three tiers from the Reel — an 8 GB laptop, a 24 GB gaming GPU, and your own servers — with the official model cards, the benchmark numbers shown on screen, and the tools that run them.

Match the model to the hardware you already own, install Ollama or LM Studio, pull the official weights, and only think about H100s if you genuinely need frontier scores on your own metal.

Use this safely

A practical starting checklist

8 GB RAM laptop, no GPU needed: run Qwen 3.5 4B (MMLU-Pro 79.1, GPQA Diamond 76.2, LiveCodeBench 55.8) or Gemma 4 E4B (69.4 / 58.6 / 52.0). Qwen is the stronger all-rounder; Gemma adds image and audio input.
24 GB VRAM card such as an RTX 4090 (an RTX 4080 works with a smaller quantisation): run Qwen 3.5 27B (86.1 / 85.5 / 80.7) or Gemma 4 31B (85.2 / 84.3 / 80.0) for frontier-class scores at home.
Your own servers: Kimi K3 is a 2.8-trillion-parameter MoE (104B active, 1M-token context) scoring GPQA Diamond 93.5, BrowseComp 91.2, Terminal-Bench 88.3 and SWE-bench Verified 76.8. Plan on six to eight H100s.
Setup: install Ollama and run "ollama run qwen3.5" or "ollama run gemma4", or use LM Studio's GUI. Pick a quantisation that fits your RAM or VRAM with headroom, and keep the model card's recommended sampling settings.
The price tag: one H100 80 GB is $25K–$40K, an 8× HGX H100 server $250K–$320K, and cloud rental $2–7 per GPU-hour, versus roughly $20 a month for a hosted subscription. Run small models locally for free; rent GPUs only for bursts.

Source links

Open the source, not a summary

Qwen 3.5 4B model card
The 8 GB laptop pick.
Qwen 3.5 27B model card
The 24 GB VRAM pick; 262K-token context natively.
Gemma 4 E4B (instruction-tuned) model card
Text, image, and audio input on a laptop.
Gemma 4 31B (instruction-tuned) model card
Kimi K3 model card
2.8T total / 104B activated MoE; the server-class option.
Ollama: Qwen 3.5
Ollama: Gemma 4
Ollama: Kimi K3
Download Ollama
LM Studio
llama.cpp on GitHub