SAICR Fair Model Use Licenses
Licenses used for models released by SAICR. Every SAICR model states in its model card which of these licenses applies to it.
Current version: 3.0 NC NF
saicr-fair-model-use-license-3.0-nc-nf.license
is the current Non-Commercial, No Fine-Tuning Except Published Scientific
Research license. Commercial use requires SAICR's written permission.
- Research modifications: Fine-tuning, including LoRA, QLoRA, adapters, and continued pretraining, is allowed solely for non-commercial scientific research. Results, including negative or inconclusive results, must be publicly and freely published within 12 months of the first modification in the project. Unpublished modifications are prohibited.
- Modified weights: Every derivative remains under the exact 3.0 NC NF license. Publications and distributed or hosted derivatives must attribute SAICR, name and link to the original model, and describe the modifications.
- Other modifications: The existing bans on pruning, merging, and distillation remain. In particular, research does not permit distillation into different architectures, merging with models under different licenses, or pruning for deployment.
- Activation steering: Amplifying or sustaining negative internal-state directions beyond training-data magnitudes, deploying instances with sustained negative-valence steering, publishing tools or benchmarks for escalating negative steering, and transferring extracted directions to steer negative states in other models are prohibited. These restrictions apply regardless of research purpose; interpretability permissions do not authorize activation steering.
- Research disclosures: At or above the current 50 million total parameters threshold, the public research results must include parameter counts by component, data source names and approximate token counts, total training tokens, hardware, duration, final loss, at least three standard public benchmarks, M ablations with and without the evaluation stream, factual observations of emergent behavior and persistent unit preferences or specialization, architecture changes from SAICR/NACR, and monitored feature summary statistics if monitoring was used. The data itself and raw activations are not required.
- Dynamic threshold: SAICR may lower the threshold based on published evidence of emergent behavior at smaller scales. Revisions must be announced publicly here with supporting evidence and an effective date. The threshold will never be raised above 50 million parameters.
The full license defines the publication deadline, parameter counting, disclosure scope, and treatment of components that are absent.
Older versions: 2.0
All existing 2.0 license files are preserved unchanged. Publishing 3.0 does not automatically relicense models released under 2.0; the license named in each model card continues to apply unless SAICR re-releases that model under a different version.
2.0 overview
| License | Commercial use | Fine-tuning / derivatives | File |
|---|---|---|---|
| 2.0 NC NF | Only with permission | Not allowed | saicr-fair-model-use-license-2.0-nc-nf.license |
| 2.0 NC | Only with permission | Allowed (non-commercial, same license) | saicr-fair-model-use-license-2.0-nc.license |
| 2.0 NF | See note below | See note below | saicr-fair-model-use-license-2.0-nf.license |
NC = Non-Commercial, NF = No Fine-Tuning.
Existing 2.0 NF file discrepancy: The file named
saicr-fair-model-use-license-2.0-nf.license currently contains the same
NC license text and identifier as the 2.0 NC file. Its filename does not
match its contents. It is retained unchanged as an older file; consult
the full text and contact SAICR to resolve the applicable license.
What the three 2.0 variants have in common
- Research, academic use, personal use, inference, and study are always allowed.
- Quantization (GGUF, GPTQ, AWQ, EXL2, MLX, FP16/INT8/INT4, ...) is always allowed.
- Outputs generated by the model can be used for any lawful purpose, including commercial purposes.
- Redistribution must keep the exact same license, with a full copy included.
- Hosted services (APIs, demos, chat interfaces) must name the model and its license. SAICR may request that a hosted service be taken down.
2.0 short summaries
2.0 NC NF — Non-Commercial, No Fine-Tuning
The most restrictive license. You may run and study the model, but not use it commercially without permission. No fine-tuning (including LoRA and adapters), pruning, merging, or distillation, and outputs may not be used to train other models. Free public APIs are allowed.
2.0 NC — Non-Commercial
Fine-tuning, LoRA, pruning, merging, and distillation are allowed for non-commercial purposes. Every derivative, including models trained on outputs, must be released under this same license, name the original model, and must not be presented as an official SAICR model. Free public APIs are allowed.
2.0 NF — No Fine-Tuning
The NF label means No Fine-Tuning, but the existing file contains the NC license text. See the discrepancy note above; this filename alone does not establish permission for commercial use.
Also include a full copy of the license file in the model repo itself, since redistribution requires one.
Permissions and contact
Requests for commercial use (NC licenses) or for fine-tuning and other modifications (NF licenses) can be sent to banaxitech@gmail.com.
This README is a summary for convenience only. The license files are the binding terms; if anything here differs from a license file, the license file applies.