Topic · 6 resources
Open Source AI
Open-weight models you can run, fine-tune and self-host: what they cost, where they match the closed alternatives, and where they genuinely do not.
"Open source" is doing a lot of work in this corner of AI, and mostly inaccurately. Almost every model described that way is open *weights*: you can download the parameters, run them, and fine-tune them, but the training data and the process that produced them are not published. That is meaningfully different from open-source software, and the licences attached range from genuinely permissive to ones with commercial restrictions that will matter to you.
Read the licence before you build on one. It is the step teams skip and the one that is expensive to discover late.
The real reason to care is control rather than cost. Running a model yourself means your data never leaves your infrastructure, nobody deprecates the version you built on, and you can fine-tune in ways a hosted API will not permit. For regulated industries and anyone with a hard data-residency requirement, that is not a preference — it is the only option that meets the constraint.
Cost is more nuanced than the headline suggests. Per-token API pricing wins until your volume is high enough that dedicated hardware is cheaper, and that crossover point is further away than most people estimate, because it has to cover the serving stack, the scaling, the monitoring and the engineering time to keep all of it running. Below sustained high volume, a hosted API is usually cheaper once you count honestly.
Where open models genuinely compete is the middle of the market. A mid-sized open model fine-tuned on your domain frequently beats a larger general model on your specific task and costs less to run — which is a better argument for self-hosting than chasing the largest checkpoint you can fit. Where they still trail is the hardest reasoning, and the ecosystem around them: fewer integrations, less tooling, fewer documented prompt patterns.
Courses1
Glossarys3
Parameters
Parameters are the learned numerical weights inside a model, adjusted during training. Parameter count — 7B, 70B, 405B — is a rough proxy for capacity, but not a reliable predictor of quality.
Open-weights model
An open-weights model has publicly downloadable parameters, letting anyone run it on their own hardware. Llama, Mistral, Qwen and DeepSeek are prominent examples.
Mixture of experts (MoE)
Mixture of experts is an architecture where only a subset of the model activates per token. A router selects a few specialised sub-networks, so a very large model runs at the cost of a much smaller one.
Prompts2
Structured Output (JSON & Tables) — Hugging Face (open-source)
When you need output your app or spreadsheet can actually use, explicitly request a schema.
Structuring Image Prompts
Good image prompts (for DALL·E, Stable Diffusion / Hugging Face models, Midjourney-style tools) generally follow: Subject -> Setting -> Style/medium -> Lighting/mood -> Composition/technical details..