Last reviewed: May 25, 2026

What is an open-source AI model? Definition and right instinct vs wrong instinct

An open-source AI model is a foundation model whose weights and architecture are freely downloadable and usable under a permissive licence (Apache 2.0, MIT). It contrasts with the proprietary model (Claude, GPT, Gemini) accessible only via API. The choice engages sovereignty, cost, and long-term flexibility.

The open-source model landscape transformed between 2023 and 2026. Proprietary models (Anthropic's Claude, OpenAI's GPT, Google's Gemini) keep a slight lead on the hardest benchmarks, but the gap is narrowing rapidly. The dominant open-source models in May 2026 are DeepSeek V4 Pro (MIT, China), Llama 4 Scout and Maverick (Meta custom licence, 10M-token context), Qwen 3.5 (Apache 2.0, Alibaba, 201 languages), Mistral Large 3 and Medium 3.5 (Apache 2.0, France), and Gemma 4 (Apache 2.0, Google). The dominant architecture is sparse MoE: all flagships have between 397 and 1,600 billion total parameters, but only 17 to 49 billion are active per token. Apache 2.0 and MIT are the most permissive licences; the Meta custom licence imposes a user-threshold clause (700M MAU) and EU restrictions.

Concrete example

A 180-employee French accounting firm evaluates in 2026 the deployment of an internal AI assistant on sensitive documents (client balance sheets, tax filings). Three options. Option A: Claude Sonnet 4.6 via Anthropic API, European Frankfurt hosting, $2,100/month, data transiting through Anthropic servers. Option B: Mistral Large 3 self-hosted on a GPU cluster rented at Scaleway, $4,800/month (4 H200 GPUs and storage), data 100% France. Option C: Mistral Medium 3.5 (24B) on 2 self-hosted A100 GPUs in the firm's datacentre, $1,200/month after 24-month hardware amortisation, complete sovereignty. The firm chooses option C for tax confidentiality and a lower TCO than the other two.

Good vs bad reflex

BAD REFLEX

Choosing a proprietary model by default, or choosing an open-source model by ideology. Both stances are symmetrical errors. The first lets significant savings slip away (a self-hosted Mistral Medium 3.5 costs 3 to 5 times less than a Claude Sonnet 4.6 at high volume) and exposes your business data to the provider's servers. The second underestimates the hidden costs of self-hosting: dedicated MLOps team, GPU amortisation, model updates, monitoring. Without sufficient volume, self-hosted open source costs more than the proprietary API.

GOOD REFLEX

Arbitrating on three operational axes and one strategic axis. Operational axes: expected inference volume (above 50 million tokens per month, self-hosting becomes profitable), data sensitivity (GDPR, professional secrecy, intellectual property often impose self-hosting), service criticality (a proprietary model can be deprecated without notice). Strategic axis: long-term independence from a single provider. The WYP rule: start with the proprietary API to learn quickly, switch to self-hosted open source on critical use cases once volume and needs are stabilised.

Sources

Open-Source LLMs Landscape 2026, Codersera, updated May 1, 2026 (compared benchmarks DeepSeek V4 Pro, Llama 4, Qwen 3.5, Mistral Large 3, Gemma 4). https://codersera.com/blog/open-source-llms-landscape-2026/ (accessed 2026-05-25)
Mistral AI official documentation, Large 3 and Medium 3.5 models, Apache 2.0, 2026. https://mistral.ai/ (accessed 2026-05-25)

← Back to glossary

What is an open-source AI model? Definition and right instinct vs wrong instinct

Concrete example

See also

Further reading

Sources