Back to Blog
12 min readAI & Open Source

Open Source in the Age of AI: The State of Open Models

In 2025 open-source LLMs closed the gap with proprietary models. In 2026 they're on par in many areas — and ahead in some. For businesses that means more control, less lock-in, and easier compliance.

The open-weight tipping point

For years, "open" models trailed the frontier by a generation. That's no longer true. The open ecosystem has matured to the point where, for most use cases, you no longer need to pay for a cloud API — the best models run for free on your own hardware. As of May 2026, DeepSeek V4 Pro tops the open-weight rankings (and leads on agentic tasks), Qwen sits right behind it, Llama 4 owns ultra-long context at 10M tokens, and a newly permissive Mistral rounds out a credible European option.

Proprietary API: best-in-class, but your data leaves, costs scale per token, and you're locked to one vendor.
Open weights: run it yourself, fine-tune freely, keep data in-house, and swap models as the leaderboard moves.

The leading open models in 2026

ModelLicenseWhy it matters
DeepSeek V4MITV4 Pro & Flash (Apr 2026), 1M-token context. Ranks #1 among open weights and #1 agentic — known for reasoning depth.
Qwen 3.5Apache 2.0Native vision-language, 397B total / 17B active params, 201 languages, 1M-token context. One of the safest enterprise picks on license alone.
Llama 4 ScoutLlama CommunityFits on a single H100 and offers an industry-leading 10M-token context window — #1 for ultra-long context.
Mistral Large 3 / Small 4Apache 2.0Now shipped under Apache 2.0, a major shift from earlier restrictive licensing. Strong European option.
GLM-5MITAnother permissive, do-anything license — fine-tune and deploy commercially with zero royalties.

Why open source wins for business

Control & data residency

Run the model on your own infrastructure and your prompts, documents and customer data never leave your network — a hard requirement for regulated industries and a strong fit for GDPR.

No vendor lock-in

Permissive licenses (Apache 2.0 on Qwen and Mistral, MIT on DeepSeek and GLM) let you fine-tune and deploy commercially with zero royalties — and switch base models without rewriting your stack.

Cost at scale

Per-token API bills grow with usage; a self-hosted open model turns inference into a fixed hardware cost. For high-volume workloads, that math flips quickly in favour of open weights.

Sovereignty & transparency

Open weights can be inspected, audited and run anywhere — important for organisations that need to understand and certify the systems they depend on rather than trust a black box.

How to choose

Start from your constraint. If license flexibility is the priority, Qwen 3/3.5 (Apache 2.0), DeepSeek (MIT) or GLM-5 (MIT) let you do essentially anything. If you need the deepest reasoning, DeepSeek leads; for documents that don't fit in a normal window, Llama 4 Scout's 10M-token context is unmatched; for a European-licensed option, Mistral now ships under Apache 2.0. The healthy default in 2026 is to build behind a thin abstraction so you can route each task to the best open model and re-evaluate every quarter.

  • ✓ Check the actual license on the model card before committing
  • ✓ Match context length and reasoning depth to your real workload, not benchmarks
  • ✓ Keep an abstraction layer so models stay swappable
  • ✓ Self-host when data residency or volume economics demand it

Want to deploy open-source AI in-house?

I help teams pick, self-host and integrate open-weight models — keeping your data private and your stack vendor-neutral.

Plan My Open-Source AI Stack →