Meta — Llama
One of the most popular open-source model families that helped make local AI mainstream.
Llama 3.2 — smaller, great for running locally
Llama 4 — larger and more capable
💡 Models like Llama 3.2 (3B) are called Small Language Models (SLMs) — the "3B" means 3 billion parameters, making them lightweight enough to run on your own machine.
Mistral AI — Mistral / Mixtral
A French AI company known for open-source innovation. Their Mixtral model uses a Mixture of Experts (MoE) approach — instead of one giant model, multiple smaller specialist models work together internally, routing requests to the right expert. Result? Better efficiency at lower compute cost. ⚡
Alibaba — Qwen
A powerful but lesser-known open-source family from Alibaba Cloud. Strong reasoning and performance — a solid alternative to Llama. 💪
Google — Gemma
Google's open-source counterpart to Gemini. Known for extremely lightweight variants — some as small as 270M parameters — making them super fast and low on resources, though less capable than larger models.
Microsoft — Phi
Microsoft's lightweight model family. Phi-4 is known for efficiency, tool calling, and being highly practical for commercial use cases. 🛠️
DeepSeek
One of the most talked-about open-source models — not because it beat GPT, but because it achieved frontier-level capability at a fraction of the training cost.
OpenAI training cost → 💸 $100M+
DeepSeek → significantly lower
The flagship model has 671 billion parameters — too large for local machines. But smaller versions are available thanks to distillation — a process where a big model teaches a smaller model, creating a compact but capable version. ✨
OpenAI — GPT OSS
Surprising but true — OpenAI released an open-weight GPT variant too! Available in multiple sizes for local use, experimentation, and free access. 🎉