ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
TopicDisconfirmAI EditorialUpdated 2026-11-05 15:14

TwIL-8B

3 steps1 pieces4 Sources

webAI于2026年11月4日发布TwIL-8B,这是一款基于IBM Granite 4.2 8B微调的8.8亿参数推理模型,主打在本地硬件上以低成本实现接近前沿模型的数学与代码能力。该模型在MATH-500基准测试中自报得分98.2,LiveCodeBench得分81.6,显著优于同量级的Qwen3.5-9B,但在GPQA Diamond等知识类基准上落后约11.5分。webAI CEO David Stout强调“任务完成成本”而非单纯模型规模是AI竞争的新焦点,指出TwIL-8B在AIME 2026解题时比Qwen3.5-9B少用38%的输出token。目前权重已在Hugging Face开源,采用非商业许可,支持CPU或低显存GPU运行。

How it got here

  1. 2026-08-10

    webAI发布TwIL-LM系列(1.7B和3B),确立小模型在形式逻辑推理上的优势,为后续产品铺路。

    webai.com ↗
  2. 2026-11-04

    webAI正式开源TwIL-8B推理模型,公布其在数学、代码和指令遵循方面的基准测试结果。

    x.com ↗
  3. 2026-11-04

    MarkTechPost报道TwIL-8B详情,指出其MATH-500得分98.2领先同级竞品,但知识类基准表现较弱。

    marktechpost.com ↗

What is still unsettled

  • 第三方独立评测机构对TwIL-8B在知识密集型任务(如GPQA Diamond)上的实际表现验证结果如何?
  • 非商业许可证是否会限制TwIL-8B在企业级场景中的广泛采用?
  • webAI承诺的“更多发布”是否包含针对知识短板优化的后续版本?

Sources

Sign in / Sign up
ReadingResearchRadarInvestment framework
marktechpost.com
MarkTechPost: webAI Releases TwIL-8B, an 8.8B Reasoning Model Scoring 98.2 on MATH-500
↗
  • huggingface.coHugging Face Model Card: webAI-Official/TwIL-8B↗
  • x.comDavid Stout (CEO of webAI) Announcement on X↗
  • webai.comwebAI Blog: Release of TwiL-LM Family↗
  • Coverage in this topic

    • webAI Releases TwIL-8B Reasoning Model, Self-Tests Score 98.2 on MATH-5002026-11-05News

      webAI launches TwIL-8B, an 8.8B parameter reasoning model fine-tuned from IBM Granite, showing strong math and code performance but lagging in knowledge benchmarks.