Summary:

webAI's November 3 checkpoint update is available to download. The current license restricts use to research and education, while evaluation results need variant-by-variant scrutiny.

webAI has added a shorter-reasoning balanced checkpoint to TwIL-8B, giving researchers another way to run its compact reasoning model locally. The weights are available, but the current license limits use to noncommercial research and education.

The useful news is the checkpoint update, rather than a first-ever launch. The official repository history records an earlier model upload on October 14, followed by length-compressed weights and the balanced/verbose arrangement on November 3. The license changed on November 4.

What is available

The model card describes an 8.79-billion-parameter model derived from IBM Granite 4.2 8B. Balanced is the default; verbose remains a separate download. The additional training behind balanced targets shorter reasoning traces. Both full-precision and quantized files are listed.

For teams evaluating local models, the distinction matters: a smaller download does not establish low end-to-end memory use or fast answers. Long reasoning traces still consume compute. Quantized files also should not be assumed to reproduce scores measured on full-precision weights.

Read the benchmark labels before comparing

webAI reports results across mathematics, coding, knowledge and instruction following. Its comparison includes Qwen3.5-9B, Gemma 4 12B and Gemma 4 31B, alongside older frontier-model versions. These are developer-reported TwIL results beside published competitor figures, not an independent, uniformly rerun evaluation.

Several rows use the verbose checkpoint and have not been evaluated on balanced. The card also leaves TwIL's LiveCodeBench version and test window unspecified, while its HLE scoring differs from the competitors' scoring. Those limitations prevent a reliable numerical ranking across the cited models. PromptsZone is therefore not reproducing a scorecard that could imply matched conditions or transfer verbose results to the new default.

A useful independent comparison would fix the exact checkpoint, benchmark release, prompts, reasoning budget, scoring method and precision for every entrant. Report accuracy together with latency and resource use. Without that common setup, a higher published number can reflect an evaluation difference as well as a model difference.

The license is narrower than the base model's

The current webAI Non-Commercial License ver. 1.0 limits the work and derivatives to noncommercial research and educational purposes. It also imposes redistribution and attribution conditions. The repository separately retains IBM's Apache 2.0 license for the base model; that does not make webAI's checkpoint commercially permissive.

Businesses should resolve licensing before planning a production deployment. Downloadable weights provide an evaluation route, not an automatic commercial-use grant.

A practical evaluation plan

  • Pin the repository revision and record whether the test uses balanced or verbose.
  • Use tasks representative of the intended application, with a fixed reasoning budget.
  • Measure answer quality, time to completion and memory on the actual target hardware.
  • Test failure handling separately before considering any user-facing deployment.

The checkpoint is an opportunity for a controlled local experiment. Its value will depend on results under those conditions, and on whether the intended use fits the license.

Image: original PromptsZone explanatory diagram, not a product screenshot. The unchanged official webAI provider logo is used as a fallback under the press kit's editorial-use guidance; no endorsement implied.