New · France 2027 presidential election: what the candidates propose on AI, quoted and sourced. Explore the tracker →

Last reviewed:

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 more or less permissive licence (Apache 2.0, MIT or a custom licence). It contrasts with the proprietary model (Claude, GPT, Gemini), accessible only via API or app. The choice engages sovereignty, cost, and long-term flexibility.

This page in Markdown

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 main open-weight models in October 2026 are DeepSeek V4 Pro (MIT, China, final release August 2026), Qwen 3.8 (Alibaba, August 2026: Qwen3.8-Max licence for the flagship, Apache 2.0 for the 27-billion-parameter version), Mistral Large 3 (Apache 2.0, France), Mistral Medium 3.5 (modified MIT licence, France), Gemma 4 (Apache 2.0, Google) and gpt-oss (Apache 2.0, OpenAI). Llama 4 Scout and Maverick (April 2025) remain Meta's latest open weights; in April 2026 Meta launched Muse Spark, whose weights are not published. The dominant architecture among flagship models is sparse MoE: from 675 billion (Mistral Large 3) to 2,400 billion (Qwen 3.8) total parameters, of which only 41 to 95 billion are active per token. Mistral Medium 3.5 is an exception: it is a dense model with 128 billion parameters. Apache 2.0 and MIT are the most permissive licences; the Llama 4 licence sets a 700-million monthly-active-user threshold and does not grant rights to its multimodal models, including Scout and Maverick, to companies with their principal place of business in the EU; the Qwen3.8-Max licence requires a separate licence for model-as-a-service or AI work assistant providers above $50 million in revenue over 12 months; the modified MIT licence of Mistral Medium 3.5 excludes companies whose consolidated monthly revenue exceeds 20 million dollars.

Concrete example

Illustrative case. 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 5.5 via the Anthropic API ($2 per million input tokens, $10 per million output tokens), data processed on Anthropic's infrastructure (the API offers no Europe-only processing, only "global" or "US"). Option B: Mistral Large 3 self-hosted on GPUs rented from a French host, data 100% in France; with 675 billion parameters, it requires a high-end GPU server. Option C: Mistral Medium 3.5 (a dense model with 128 billion parameters) on a server with 8 H100- or H200-class GPUs installed in the firm's datacentre, the reference setup published by Mistral. The modified MIT licence is not an issue: the firm is very far below the $20M monthly revenue threshold. Heavy hardware investment, complete sovereignty. The firm chooses option C for tax confidentiality, after checking that it has the skills to operate this server.

FAQ

What is an open-source AI model?

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

Are open-source models as capable as proprietary ones?

Proprietary models keep a slight lead on the hardest benchmarks, but the gap is narrowing fast. The landscape transformed between 2023 and 2026: in October 2026, the main open-weight models are DeepSeek V4 Pro, Qwen 3.8, Mistral Large 3, Gemma 4 and gpt-oss.

Open source or proprietary: how do we decide?

By avoiding the two symmetrical errors, choosing proprietary by default or open source by ideology. The decision rests on three operational axes (expected inference volume, data sensitivity, service criticality) and one strategic axis (long-term independence from a single provider). The WYP rule: start on a proprietary API to learn fast, then switch to self-hosted open source on critical use cases once volume and needs have stabilised.

When does self-hosting become profitable?

There is no universal threshold. Work out the break-even point by comparing the monthly API bill with the full cost of a server: GPUs to amortise, energy, an MLOps team, model updates and monitoring. At current API prices ($2 per million input tokens for Claude Sonnet 5.5), it lies well beyond a few tens of millions of tokens per month. Self-hosting is justified first by confidentiality or sovereignty, rarely by cost.

Should we be wary of so-called open-source licences?

Yes, not all are equally permissive. Apache 2.0 and MIT largely are, but Meta's licence for Llama 4 imposes a threshold of 700 million monthly active users and EU restrictions, and the modified MIT licence of Mistral Medium 3.5 excludes companies whose consolidated monthly revenue exceeds 20 million dollars. The licence must be checked before any deployment.

See also

Further reading

Mistral Medium 3.5, model card, Mistral AI (Hugging Face) (external resource)

Sources

  1. Mistral Medium 3.5 128B, model card, Mistral AI, Hugging Face (dense 128-billion-parameter model, modified MIT licence: excluded above $20M monthly revenue, vLLM deployment on 8 GPUs). https://huggingface.co/mistralai/Mistral-Medium-3.5-128B (accessed 2026-09-29)
  2. Mistral Medium 3.5, official Mistral AI documentation (released 28 April 2026, open weights, modified MIT licence). https://docs.mistral.ai/models/mistral-medium-3-5-26-04 (accessed 2026-09-29)
  3. Mistral Large 3, model card, Mistral AI, Hugging Face (MoE, 675 billion parameters with 41 billion active, Apache 2.0). https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512 (accessed 2026-09-29)
  4. DeepSeek-V4-Pro, model card, DeepSeek-AI, Hugging Face (MoE, 1.6 trillion parameters with 49 billion active, MIT licence). https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro (accessed 2026-09-29)
  5. Qwen3.5-397B-A17B, model card, Alibaba Qwen, Hugging Face (397 billion parameters with 17 billion active, 201 languages, Apache 2.0). https://huggingface.co/Qwen/Qwen3.5-397B-A17B (accessed 2026-09-29)
  6. Gemma 4 31B, model card, Google DeepMind, Hugging Face (Apache 2.0). https://huggingface.co/google/gemma-4-31B-it (accessed 2026-09-29)
  7. The Llama 4 herd, Meta AI, April 2025 (Scout: 10-million-token context; Maverick: 400 billion parameters with 17 billion active). https://ai.meta.com/blog/llama-4-multimodal-intelligence/ (accessed 2026-09-29)
  8. Llama 4 Community License and acceptable use policy, Meta (700 million monthly active users threshold, EU restriction on multimodal models). https://github.com/meta-llama/llama-models/tree/main/models/llama4 (accessed 2026-09-29)
  9. Data residency, Claude Platform documentation, Anthropic (inference_geo: "global" or "us"). https://platform.claude.com/docs/en/manage-claude/data-residency (accessed 2026-09-29)
  10. Qwen3.8-2.4T-A95B, model card and Qwen3.8-Max licence, Alibaba Qwen, Hugging Face. https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B (accessed 2026-10-05)
  11. DeepSeek-V4-Pro-0813, model card (MIT licence), DeepSeek-AI, Hugging Face. https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813 (accessed 2026-10-05)
  12. gpt-oss-120b, model card (Apache 2.0), OpenAI, Hugging Face. https://huggingface.co/openai/gpt-oss-120b (accessed 2026-10-05)
  13. Introducing Muse Spark, Meta AI, 8 April 2026. https://ai.meta.com/blog/introducing-muse-spark-msl/ (accessed 2026-10-05)

Editorial methodCorrections log

← Back to glossary

Address copied