A model page gives you a benchmark chart, a download button, and somewhere below them a license. The Openness Profile puts another set of questions beside those familiar claims: what can you obtain, what may you do with it, and how much of the work can another person check?
The profile has a tier label and six separate dimensions. It does not add them into a score out of 100. Access, legal permission, and the possibility of rebuilding a model are different properties; an average can hide the missing one. The website shows the profile and each dimension’s rationale alongside its evidence.1
The unit is a specified release or a group of releases with equivalent relevant terms. A family whose checkpoints use incompatible licenses can have an unscored overview linking the release profiles. A single score for that family would pretend away the distinction a user needs.
The six questions
The rubric and its structured data live with the companion dataset. Scores run from 0 to 3: No, Limited / partial, Yes with caveats, and Strong / exemplary.2
Access asks whether weights and inference artifacts are downloadable. An announcement about future publication is not the file.
Permission asks whether users can use, modify, and redistribute the artifacts for their purpose. The assessment records the artifact license and any additional obligations accepted to obtain it. A permissive copyright license and a restrictive access contract can coexist.
Deployment asks whether the weights and runtime can be operated independently on chosen hardware. Downloading an artifact and being able to run a usable system are separate achievements.
Transparency asks whether the publisher documents training data, processing, post-training, alignment, and evaluation. Documentation is evidence to inspect, not a guarantee that its account is complete or correct.
Reproducibility asks whether a skilled group could build a substantially equivalent system from what was published, given time and compute. It does not require bit-for-bit identity, and the score does not certify that an independent group has completed the reconstruction.
Governance & Participation asks what outsiders can contribute or decide without employment or affiliation with the publisher. Its labels run from No participation path, through Cosmetic participation and Real but bounded participation, to Genuine community governance.3 A support forum demonstrates a contact path. Merged changes demonstrate accepted work. Authority over a release requires evidence about decision rights as well.
People assign these dimension scores by interpreting sources. The rationales therefore matter as much as the numbers. If a reviewer cannot explain why the evidence earns 2 rather than 1, the score needs work. A precise tier calculation cannot cure an unsupported input.
How the tier is assigned
The tier is calculated from Access, Permission, Deployment, Transparency, and Reproducibility. Governance remains visible as a separate axis. An independently runnable, reproducible model can still have centralized governance; a participatory community can also lack usable weights.
The calculation follows the tests below in order. A, P, D, T, and R abbreviate those five dimensions.4
| First matching test | Tier | Meaning |
|---|---|---|
| A = 0 and D = 0 | 0: Proprietary service | No downloadable weights and independent runtime in the assessed release. |
| Otherwise, A < 3 or D < 2 | 1: Access-limited | Artifacts or independent deployment remain substantially limited. |
| Otherwise, P < 2 or D < 3 | 2: Restricted or deployment-limited | Weights are accessible, with material permission or deployment limits. |
| Otherwise, A, P, D, T, and R are all 3 | 5: Reproducible open release | Strong evidence across the five release dimensions, including Permission 3. |
| Otherwise, T and R are both at least 2 | 4: Documented research release | Meaningful training documentation and reconstruction resources, with remaining caveats. |
| Otherwise | 3: Practical open-weight | Strong practical access and deployment, with less complete training evidence. |
These rules give identical five-dimension inputs identical tiers. A changed score can change the result; an unexplained publisher preference cannot. The validator checks both the score ranges and the resulting tier, and each scored entry records the reason for its placement.
No tier independently certifies compliance with the OSI Open Source AI Definition. That definition has its own requirements. Tier 4 can include a release with a use-restricted license or an additional contract; Tier 5 is still an assessment based on the available evidence, not an external certification. The profile helps locate questions. It does not replace reading the underlying standard or terms.
Different profiles, explainable differences
Three examples show why a tier needs its dimensions underneath it. These are the assessed release profiles, rather than judgments about everything from each publisher. In the score column, the five numbers are A/P/D/T/R in that order.5
| Release | A/P/D/T/R | Tier | Main reason |
|---|---|---|---|
| Gemma 4 | 3/3/3/1/1 | 3 | Apache weights; corpus and full recipe unpublished. |
| StarCoder2 | 3/2/3/3/2 | 4 | Training resources; use restrictions and reconstruction caveats. |
| Apertus 1.5 70B | 3/2/3/3/3 | 4 | Extensive resources; additional AUP obligations. |
Gemma 4 and StarCoder2 make the tradeoff especially clear. Google gives Gemma 4 Apache 2.0 terms and publishes model documentation, but not the corpus and full training recipe. Its official community forum provides a participation channel; it does not give outsiders control of upstream model development.678
StarCoder2 uses BigCode OpenRAIL-M, with restrictions that make it different from an OSI-approved software license. It also publishes training-data resources, opt-out handling, code, and evaluations.910 The profile records both facts. More training evidence does not erase the permission caveat, and a cleaner license does not create missing data.
Apertus’s extra AUP is likewise visible in Permission. The underlying Apache label is recorded separately from the download agreement’s indemnity and downstream assent obligations. Its inputs retain their own dataset terms.11 With P at 2, it cannot qualify for Tier 5 under these rules, even if all the other release scores are 3.
Governance is shown beside these differences, without changing their tiers. Evidence should identify what outsiders can actually influence. A count of outside pull requests to an evaluation repository is useful evidence of contribution, but cannot alone establish shared authority over the model release.
Reading the profile over time
Every profile needs a scope, a dated dataset snapshot, sources, and rationales. The snapshot date describes that published assessment; it should not be mistaken for proof that every cited page was rechecked that day. License revisions and newly released artifacts may require a new assessment.
The website and book draw from the same maintained data, while a printed edition necessarily describes a frozen point in time. A later site update can legitimately differ from the book. Readers should be able to retrieve the edition’s snapshot and distinguish a correction from a subsequent release change.
For a decision, read the tier, the weak dimensions, and the evidence behind them. Then verify the exact checkpoint and terms you will use. The profile has done its job when it helps you find the question that a download button left unanswered.
Notes
But Is It Open?, living index and release profiles, https://butisitopen.com/is-it-open/; edition data snapshot, November 15, 2026, https://github.com/butisitopen/data/tree/main/snapshots/2026-11-15. The living site can change; the dated snapshot records the assessments used here. ↩
But Is It Open? companion dataset, November 15, 2026 snapshot, https://github.com/butisitopen/data/tree/main/snapshots/2026-11-15, rubric dimensions and dimension scale, with the accompanying JSON Schema. Scores are qualitative assessments supported by sources and rationales, distinct from the tier calculated from them. ↩
But Is It Open? companion dataset, governance scale, November 15, 2026 snapshot, https://github.com/butisitopen/data/tree/main/snapshots/2026-11-15. Participation and governance require evidence about the role outsiders can hold, rather than an inference from the artifact license. ↩
But Is It Open? companion dataset, tier rules and release rationales, November 15, 2026 snapshot, https://github.com/butisitopen/data/tree/main/snapshots/2026-11-15. The ordered rules use A, P, D, T, and R; G is excluded. Tier 5 requires P = 3. A scored entry must match the calculated tier rather than merely pass a lower tier’s minimum checks. ↩
But Is It Open? companion dataset, November 15, 2026 snapshot, https://github.com/butisitopen/data/tree/main/snapshots/2026-11-15, Gemma 4, StarCoder2, and Apertus 1.5 70B release profiles. The table reproduces the five release-dimension scores and calculated tiers. These labels are not OSI certification. ↩
Google DeepMind, “Gemma 4 model card,” Google AI for Developers, last updated July 30, 2026, https://ai.google.dev/gemma/docs/core/model_card_4. The card lists five sizes (E2B, E4B, 12B, 26B A4B, 31B) in pre-trained and instruction-tuned variants, deployable “from high-end phones to laptops and servers.” Technical report: Gemma Team, “Gemma 4 Technical Report,” arXiv:2607.02770, https://arxiv.org/abs/2607.02770. ↩
The Gemma 4 model card gives the license as Apache 2.0 and links Google’s copy of the text: “Apache License 2.0,” Gemma, Google AI for Developers, https://ai.google.dev/gemma/apache_2; canonical text at https://www.apache.org/licenses/LICENSE-2.0. Launch post: Clement Farabet and Olivier Lacombe, “Gemma 4: Byte for byte, the most capable open models,” Google blog, April 2, 2026, https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4. Earlier Gemma generations used custom terms, https://ai.google.dev/gemma/terms. ↩
Google AI Developers Forum, “About the Gemma Forum,” May 9, 2024, https://discuss.ai.google.dev/t/about-the-gemma-forum/1506; Gemma discussions, https://discuss.ai.google.dev/c/gemma/10. These establish a public participation channel, not community authority over upstream model releases. ↩
BigCode, “BigCode Model License Agreement” (BigCode OpenRAIL-M v1), Hugging Face, https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement; the StarCoder2 model card names this license, https://huggingface.co/bigcode/starcoder2-7b. The dataset entry
starcoder2recordsosi_approved: falseand lists the permitted-use and prohibited-use restrictions. ↩Anton Lozhkov et al., “StarCoder 2 and The Stack v2: The Next Generation,” arXiv:2402.19173, February 29, 2024, https://arxiv.org/abs/2402.19173; training data at https://huggingface.co/datasets/bigcode/the-stack-v2-train; code at https://github.com/bigcode-project/starcoder2 (Apache-2.0), which points to the BigCode Evaluation Harness, https://github.com/bigcode-project/bigcode-evaluation-harness. ↩
Swiss AI Initiative, Apertus 1.5 70B model page and download agreement, https://huggingface.co/swiss-ai/Apertus-v1.5-70B; “Apertus LLM Acceptable Use Policy,” v1.5, July 14, 2026, https://github.com/swiss-ai/apertus-legal/blob/main/apertus_1.5/USAGE_POLICY.pdf; pretraining-data reconstruction resources, https://github.com/swiss-ai/pretrain-data. The AUP’s obligations and the input datasets’ terms are separate from the Apache license on released project code and weights. ↩
From But Is It Open? by Timothy O’Brien (Discursive, 2026). © 2026 Timothy O’Brien. All rights reserved.