This observation covers the period from 7:55 am on 8 November to 7:55 am on 9 November 2026 in Australia/Sydney and includes one verified technical change.
TwIL-8B progresses from announcement to public weights
webAI created the TwIL-8B repository on Hugging Face on 8 November and published the complete deliverables in its first commit. The previously verifiable public state was the model announcement on 4 November. The new fact in this window is the repository's release of the model card, licence, configuration, tokenizer, safetensors weights for the balanced and verbose variants, and several GGUF quantisations. The commit history contains a single root commit and the file tree directly exposes the weights and quantised files, providing primary evidence that the weights are available.
The model card describes TwIL-8B as an 8.79 billion parameter dense text model post-trained from IBM Granite 4.2 8B. Its training sequence comprises full-parameter supervised fine-tuning, GRPO/DAPO reinforcement learning for the verbose variant, and length-aware reinforcement learning for the balanced variant. Both variants use the same architecture, tokenizer and chat template. The repository root loads the balanced variant by default, while the verbose variant is in the verbose subdirectory. The 131,072-token context window is inherited from the base model, and input and output are English text. Each bfloat16 variant is about 16.38 GiB, while the balanced Q4_K_M GGUF is about 4.98 GiB and can be used for local inference.
The access conditions have a defined boundary. TwIL-8B is distributed under the webAI Non-Commercial License 1.0, while the IBM base-model material retains its Apache 2.0 licence. It is therefore a downloadable-weights release without permission for commercial use. The model card also states that no additional safety or preference tuning was applied beyond the measures inherited from Granite 4.2, and lists knowledge-heavy tasks, English-only operation and long reasoning outputs among its limitations.
The developer reports a MATH-500 score of 98.2 for the balanced variant and a LiveCodeBench score of 81.6 for the verbose variant. The model card says some marked evaluations have not yet been run on the balanced variant and that comparison-model scores were taken from published sources rather than rerun under one evaluation setup. An independent tracking page continued to label these results as self-reported at the time of this run, with no independent leaderboard result or public reproduction available. This observation therefore records the delivery of the weights, model structure, licence and disclosed evaluation status without treating the scores as independently confirmed capability results.
Sources
webAI / Hugging Face, TwIL-8B model card, 8 November 2026, https://huggingface.co/webAI-Official/TwIL-8B
webAI / Hugging Face, Publish TwIL-8B commit, 8 November 2026, https://huggingface.co/webAI-Official/TwIL-8B/commit/9563ba7dfb70eeca4a26654b195011cbf818e7d9
isitsota, Is TwIL-8B state of the art? Pending benchmarks, 4 November 2026, https://isitsota.com/m/twil-8b
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.