Weights are one part of openness
A downloadable checkpoint does not tell you whether training data, training code, and the full development process are available.
THE AI LANDSCAPE 03 / ACCESS & CONTROL
Look beyond the label. Compare downloadable weights, model licenses, and deployment choices across selected releases.
THE BIG PICTURE
Selected coverage. Check each entry for context and evidence.
Try a different name or a broader category.
A manually reviewed snapshot as of September 23, 2026. Colors organize the evidence; they do not score quality. On smaller screens, swipe across the comparison.
CONTEXT MAKES THE DIFFERENCE
Our interpretation of the comparison.
A downloadable checkpoint does not tell you whether training data, training code, and the full development process are available.
A developer can release both hosted models and downloadable models, or use different licenses within one family.
Self-hosting gives you deployment choices. It also means arranging hardware, serving software, monitoring, and updates.
TRANSPARENT BY DESIGN
Official announcements, product documentation, and model repositories. All linked sources were checked on September 23, 2026. Publication dates and review dates are shown separately.
This page compares distribution and the licenses linked by the named official repositories. “Permissive” groups checkpoints labeled Apache 2.0 or MIT; “Custom” groups developer-specific weight licenses; “Hosted only” means no downloadable weights are offered in the cited release. These categories are not a certification of open-source AI. License summaries are deliberately limited: read the linked terms for your exact version and use case. Older available checkpoints remain useful comparisons and are not presented as new releases.
Dates on continuously updated documentation are labeled “Current documentation” unless a specific publication date is available. This page is maintained manually; the review date is not a live-feed timestamp. Company announcements describe the company’s claims and plans, not an independent audit.
DeepSeek / Permissive license
The official model card labels the release MIT. Its API is one access route; downloading the weights is a separate deployment choice.
Qwen / Alibaba / Permissive license
The 27B checkpoint uses Apache 2.0. This classification belongs to this checkpoint, not to every model carrying the Qwen name.
Google DeepMind / Permissive license
Google released Gemma 4 in E2B, E4B, 26B MoE, and 31B Dense sizes under Apache 2.0. The April 2026 release remains a useful on-device and server comparison.
OpenAI / Permissive license
OpenAI released these reasoning checkpoints in August 2025. They are distinct from the hosted GPT-6 products and can be deployed on infrastructure chosen by the user.
Qwen / Alibaba / Custom license
The official FP8 model card links Qwen Community 1.0, rather than Apache 2.0. Review its commercial service conditions before treating it like the 27B checkpoint.
Meta / Custom license
Scout uses Meta’s Llama 4 Community License. Weight availability does not remove the agreement’s requirements or make this an Apache-licensed release.
OpenAI / Hosted only
The Sol and Luna announcement distributes these models through OpenAI services. It does not offer model weights for self-hosting.
Anthropic / Hosted only
Access through multiple cloud platforms provides more service choices. It does not make Opus 5.5 a downloadable open-weight model.