Post by ServiceNow AI Research

🚀 𝗔𝗽𝗿𝗶𝗲𝗹-𝟭.𝟱-𝟭𝟱𝗕-𝗧𝗵𝗶𝗻𝗸𝗲𝗿 — 𝗦𝗺𝗮𝗹𝗹 model. 𝗕𝗶𝗴 reasoning 🔥 Introducing ServiceNow’s 15B-parameter model that matches 𝗗𝗲𝗲𝗽𝗦𝗲𝗲𝗸–𝗥𝟭–𝟬𝟱𝟮𝟴, 𝗠𝗶𝘀𝘁𝗿𝗮𝗹–𝗺𝗲𝗱𝗶𝘂𝗺–𝟭.𝟮 and 𝗚𝗲𝗺𝗶𝗻𝗶 𝗙𝗹𝗮𝘀𝗵 𝟮.𝟱 on the Artificial Analysis Index (𝗔𝗔𝗜 𝟱𝟮) — delivering comparable results at a 𝗳𝗿𝗮𝗰𝘁𝗶𝗼𝗻 𝗼𝗳 𝘁𝗵𝗲 𝘀𝗶𝘇𝗲 (at least 8-10 times smaller) >> Why it matters 💡 ⚡ 𝗙𝗿𝗼𝗻𝘁𝗶𝗲𝗿-𝗹𝗲𝘃𝗲𝗹 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 on a single GPU 🧠 𝗡𝗼 𝗥𝗟 𝗽𝗵𝗮𝘀𝗲 — the step-change comes from mid-training 💬 𝗥𝗲𝗮𝘀𝗼𝗻𝘀 𝗼𝘃𝗲𝗿 𝗶𝗺𝗮𝗴𝗲𝘀 - Image + Text mid training enables model to reason over images without additional training 🧩 𝗚𝗿𝗲𝗮𝘁 𝗮𝘁 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 — AIME2025: 88, GPQA: 71, LCB: 73 ✅ 𝗙𝗼𝗹𝗹𝗼𝘄𝘀 𝗶𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗶𝗼𝗻𝘀 reliably — IFBench: 62 📡 T𝗮𝘂𝟮 𝗕𝗲𝗻𝗰𝗵 (Telecom): 68 → ready for real-world workflows 📦 𝗢𝗽𝗲𝗻 𝘄𝗲𝗶𝗴𝗵𝘁𝘀 model to further research and reproducibility (MIT license) >> Try it / get it 🔗 🧠 Model weights: https://lnkd.in/exmZC4J6 💬 Chat demo (HF Space): https://lnkd.in/eC_vx8zr ~25% of Apriel-v1.5's pre-training text data for depth up-scaling comes from NVIDIA Nemotron collection. Shoutout to NVIDIA AI for their amazing open source work. Thanks to the NVIDIA DGX Cloud team for helping debug the cluster 24*7 🔥 Grateful as well to our amazing collaborators from Turing, who helped craft the highest quality human data! Shruthan Radhakrishna, Jash Mehta, Aman Tiwari, Pulkit Pattnaik, Saloni Mittal, Kelechi O., Akintunde O., Masoud Hashemi, Rishabh Maheshwary, Shiva Krishna Reddy Malay, Aanjaneya Shukla, Oluwanifemi Bamgbose, Sathwik Tejaswi Madhusudan, Torsten Scholak, Sagar Davasam, Srinivas Sunkara, Nicolas Chapados, Raahul Srinivasan, Khalil Slimi #AI #GenAI #LLM #Huggingface #Multimodal #SOTA #ServiceNow #Reasoning #NVIDIA #DGX #OpenWeights