The models that answer the benchmark questions.
The pilot compares these two base models: the same architecture and training compute, one trained on pre-1931 text and one on modern web text. The two training sets differ in the kinds of text they contain as well as in date.
| Model | Kind | Cutoff | Size | Access | Status |
|---|---|---|---|---|---|
Talkie 1930 13B base Talkie Run in the DEPENDENCE pilot, September 2026. | Historical | Weights | piloted | ||
Talkie web 13B base Talkie Same architecture and training compute as Talkie 1930, trained on modern web text. Downloaded; not yet run. | Modern base | Weights | planned |
Also run in the pilot, for comparison and for the public site. We plan to add one deployed frontier model to this group.
| Model | Kind | Cutoff | Size | Access | Status |
|---|---|---|---|---|---|
Qwen3.5 9B Alibaba Run in the DEPENDENCE pilot, September 2026. | Open, instruction-tuned | Weights (8-bit MLX) | piloted |
Models we may add later.
| Model | Kind | Cutoff | Size | Access | Status |
|---|---|---|---|---|---|
Ranke-4B Göttlich, Loibner, Jiang, Voth | Historical | Weights | candidate | ||
OLMo 3 7B AI2 | Modern base | Weights | candidate | ||
OLMo 3 32B AI2 | Modern base | Weights | candidate | ||
Qwen3.5 9B Base Alibaba | Modern base | Weights | candidate | ||
Qwen 3.7 Plus Alibaba Run in the July 2026 pilot, now superseded. | Deployed | OpenRouter | piloted |