webAI updates TwIL-8B with shorter reasoning and noncommercial terms
webAI’s TwIL-8B update adds a shorter reasoning option. Its noncommercial license is an equally important part of the decision for anyone considering these downloadable weights.
The official repository history shows the two variants taking their current form on November 3, followed by relicensing on November 4. Earlier weights were uploaded on October 14. This is an update to an existing model, rather than its first appearance.
What changes between balanced and verbose
The 8.79 billion parameter model builds on IBM Granite 4.2 8B. Balanced adds training for shorter reasoning. Verbose uses longer reasoning traces.
| Choice | Published detail |
|---|---|
| Balanced | Default checkpoint |
| Verbose | Verbose folder |
| Q4_K_M download | 4.98 GiB per variant |
| BF16 download | 16.38 GiB per variant |
For a research comparison, keep the task, runtime and output limit fixed while changing the variant. Record whether the answer passes your checks, how long it takes and whether generation stops before completion. A shorter response is useful only when it still solves the problem.
Check the license before using the weights
The webAI Non-Commercial License version 1.0 defines permitted use as noncommercial research and educational purposes only. That restriction also applies to derivatives. Redistribution requires a complete license copy and preservation of existing notices.
The Granite base model has Apache 2.0 terms, but that does not remove the separate restrictions on these released weights. A team considering a paid service or internal commercial workflow should resolve permission before adopting this checkpoint. Download availability alone is not commercial authorization.
Keep the license file with the exact model revision in your evaluation record. This matters when a repository changes terms or replaces a checkpoint under a familiar name. A download made on a different date may require a separate review.
Read the benchmark footnotes before the headline
webAI reports 98.2 on MATH-500 and 81.6 on LiveCodeBench. The coding result is from verbose, not balanced. Competitor figures were not rerun, and TwIL-8B’s LiveCodeBench version and date window are unrecorded. The scores use BF16 weights, not the smaller GGUF downloads. These are vendor results, with incomplete comparability.
webAI says tool use and agentic tasks were not evaluated. For a repository assistant, separately test file edits, tool calls and task completion. A benchmark for generated functions is only one part of that assessment.
A useful comparison needs a declared success rule. For a generated function, that might mean passing hidden unit tests. For a constrained answer, it might mean meeting every requested field and format. Keep failed and incomplete responses in the denominator instead of showing only the strongest examples.
Plan for memory beyond the downloaded file
The quantized weights make local evaluation more approachable, but the file size is only a starting point. Leave room for the runtime, conversation state and other applications. Measure peak memory with the input length and number of simultaneous requests you expect to use.
Choose the variant explicitly in the official files list and keep its name in your test results. ByteForward has not run these checkpoints or verified their performance on consumer hardware. Our LFM3 guide covers the related process of choosing a local model by its inputs, outputs and runtime needs.
