webAI Releases TwIL-8B: An 8.8B Reasoning Model That Scores 98.2 on MATH-500
webAI has released TwIL-8B, an 8.8B reasoning model tuned for math, code, and instruction following. TwIL-8B is post-trained from IBM's Granite 4.2 8B, and its weights are available on Hugging Face under a non-commercial license.
In its launch post, webAI frames the model around cost per task rather than raw scale. The 4-bit build weighs in at 4.98 GiB, making it viable for CPU or modest-VRAM GPU inference.
TL;DR
- Size: 8.79B parameters, dense architecture. 131,072-token context window, inherited from the base model.
- Runs on: CPU or modest-VRAM GPUs via a 4.98 GiB Q4KM GGUF. Full bf16 weights are 16.38 GiB.
- Performance: Leads its size class on math and code, trails on knowledge benchmarks.
- Best result: MATH-500 at 98.2, the top score in webAI's comparison.
- Weakest result: GPQA Diamond at 70.2, 11.5 points behind Qwen3.5-9B. HLE stands at 12.6 (verbose variant).
What Is TwIL-8B?
TwIL-8B is a dense, decoder-only reasoning model from webAI. It uses 40 layers with a hidden size of 4096 and 32 attention heads, and it inherits the 131,072-token context window from Granite 4.2 8B.
As with most 2026-era small reasoning models, the release reflects a broader shift toward post-training and quantization efficiency over parameter scaling: webAI's pitch centers on delivering competitive reasoning performance at a fraction of the memory footprint typically associated with larger models.
via MarkTechPost