Open sourcewebAIConfirmed
Story 17 of 17 in the Wednesday 4 November editionwebAI releases TwIL-8B reasoning model post-trained from IBM Granite 4.2 8B
Post-trained with full-parameter SFT and GRPO/DAPO reinforcement learning in balanced and verbose modes, the 8.8B checkpoint scores 98.2 on MATH-500.
Edited by Ashutosh Shrivastava
5 Nov 2026, 09:50 IST · 1 min read
webAI released TwIL-8B on Hugging Face on 4 November 2026 under the webAI Non-Commercial License 1.0, an 8.8-billion-parameter reasoning model post-trained from IBM’s granite-4.2-8b using full-parameter supervised fine-tuning and GRPO/DAPO reinforcement learning in balanced and verbose variants, posting 98.2 on MATH-500 and 81.6 on LiveCodeBench while leading Qwen3.5-9B by 9.9 points on IFBench.
Sources
- PrimaryTwIL-8B weights and model card on Hugging Facehuggingface.co/webAI-Official
The rest of the day
Read the whole day
01CreativeOpenAI
OpenAI rolls out ChatGPT Images 2.6 with animated GIFs and two new API image models
Images 2.6 generates visuals 50% faster across ChatGPT and Codex while adding GPT-Image-2.6 Flare and Sunburst to the API with SVG export and markup-guided edits.
02ModelTencent
Tencent launches multimodal Hy4 model with native computer use and 3D engine training
Hy4 adds visual reasoning alongside CLI and GUI control, trained directly inside Blender, CAD, Unreal Engine, Unity and Godot environments.
03Open sourceMoonshot AI
Moonshot AI releases open-weight Kimi K3.1 Pro with 2.8T parameters and 1M context
Activating 104B parameters per token with native MXFP4 weights, the multimodal MoE model leads Opus 5.5 and GPT-6 Astra on SWE Marathon v1.1.
04ResearchOpenAI
OpenAI’s autonomous cyber pipeline uncovers and patches 948 Linux kernel root flaws
Powered by an upcoming cyber model, a five-stage VM validation and exploit pipeline generated working privilege-escalation exploits at $1,279 per distinct vulnerability.
Spotted an error? Email aiforsuccess.in+compass@gmail.com or tag @ai_for_success on X