Open-Source Models
Not every powerful model is locked behind an API. A whole ecosystem of open models — whose weights you can download, run, and fine-tune — now rivals the best closed models on many tasks. This article covers what "open" really means, the key players, the licences that matter, and the trade-offs between open and closed.
💡 In one line: Open models let you download and run the weights yourself — offering control, privacy, and low cost — though "open-weight" usually isn't the same as fully "open-source."
Open vs. Closed Models
| Closed / Proprietary | Open (open-weight) | |
|---|---|---|
| Access | API or app only | Download the weights |
| Examples | GPT, Claude, Gemini | Llama, DeepSeek, Qwen, Gemma, Mistral |
| Run it where? | The provider's servers | Your hardware |
| Cost model | Per token | Your compute (fixed at scale) |
| Customise | Limited | Fine-tune freely |
"Open-Weight" vs. "Open-Source" (the Key Distinction)
These terms are often mixed up, but the difference matters:
- Open-source (the ideal) — the weights, training code, training data, and a permissive licence are all available.
- Open-weight — only the weights are released; the training data and full pipeline may stay private.
Most so-called "open-source LLMs" are really open-weight (Llama, Qwen, Gemma, DeepSeek, GLM, Kimi). For a builder, the licence matters more than the label.
The Open Ecosystem
A snapshot of major open families (the lineup shifts constantly):
| Family | Maker | Note |
|---|---|---|
| Qwen | Alibaba | Versatile, multilingual; often Apache 2.0 |
| DeepSeek | DeepSeek | Strong reasoning/cost; MIT |
| Llama | Meta | Huge ecosystem; custom licence |
| Gemma | On-device friendly; Apache 2.0 | |
| Mistral | Mistral AI | European; Apache 2.0 |
| GLM / Kimi / Phi / gpt-oss | Z.ai / Moonshot / Microsoft / OpenAI | Growing open options |
Licences Matter
Before shipping a product, check the model's licence:
- Apache 2.0 / MIT — the most permissive: commercial use, modification, redistribution, no royalties.
- Custom licences (e.g. Llama's) — commercial use allowed, but with restrictions (usage caps, acceptable-use terms).
- Non-commercial (e.g. some CC-BY-NC) — research only, not for commercial products.
The "open" label alone doesn't tell you what you're allowed to do — the licence does.
Why Use Open Models?
- Data privacy & sovereignty — data never leaves your infrastructure (vital for healthcare, finance, government).
- Cost control — no per-token fees; a fixed cost at scale.
- Customisation — fine-tune on your own data.
- No vendor lock-in and more transparency.
- Offline / air-gapped deployment is possible.
The Trade-offs
Open models aren't automatically the right choice:
| Open | Closed | |
|---|---|---|
| Setup | Needs hardware + ML-ops skill | Turnkey API |
| Frontier edge | Catching up fast | Often most cutting-edge |
| Safety | Guardrails can be removed | More controlled |
| Control & cost | High control, fixed cost | Convenient, ongoing cost |
The State of Open (2026)
- The gap with closed models has narrowed dramatically — open models now win on many coding, maths, and long-context tasks.
- Mixture-of-Experts architectures dominate open flagships (efficiency at scale).
- Tools like Ollama, vLLM, LM Studio, and Hugging Face make self-hosting easy; quantisation shrinks models to run on modest hardware.
- The direction is strategic, not fixed — some labs move closed, others double down on open.
Summary
- Open models let you download and run the weights yourself, unlike closed API-only models.
- "Open-weight" ≠ "open-source" — most popular open LLMs release weights only, not training data.
- Key families include Qwen, DeepSeek, Llama, Gemma, and Mistral; licences (Apache 2.0 / MIT vs. custom) decide what you can do.
- Open models offer privacy, cost control, and customisation — at the price of hardware and ops effort.
- In 2026, open models have closed much of the gap with proprietary ones, though the open-vs-closed balance keeps shifting.