This path is for you if the names of open models have started to blur together. It separates a base model, a fine-tune, an instruct variant, and a small model you can run. No part is live on this page yet. Start with Open-source AI explained, which already covers weights, licenses, and a model card.
- 1 The Landscape: Foundations, Fine-Tunes, and “-Instruct” Map open models as foundations, fine-tunes, and instruct variants. Read the model card and license before you pin a default. Thumbnails are not architecture. Scheduled · October 25, 2026
- 2 Mistral, Gemma, Phi, and Friends (Map, Not Dump) Use a short map of Mistral, Gemma, Phi, and peers to pick roles. Avoid the release-note dump. One default beats a sticker wall. Scheduled · October 26, 2026
- 3 Code Models You Can Run Yourself: Qwen, DeepSeek, and More Treat code-tuned open models as assistants with a verify step. Keep unit tests and human review in the loop. Specialization is not a license to skip diffs. Scheduled · October 27, 2026
- 4 Small Models That Actually Run Offline Pick small quantized models for offline and edge work. Match RAM and power first. Do not expect cloud-sized prose from a pocket-sized weight file. Scheduled · October 28, 2026
- 5 Picking One Default Stack and Sticking to It Choose one default open-model stack and stick to it through a real upgrade gate. Model hopping taxes prompts, tooling, and delivery more than it buys accuracy. Scheduled · October 29, 2026
- 6 Benchmarks You Can Ignore (and the Few You Should Not) Ignore vanity benchmarks as buying guides. Keep a few high-signal public checks if you must, and invest in an internal golden harness tied to your prompts. Scheduled · October 30, 2026