Apriel-1.5-15B-Thinker: Mid-training is all you need

Shruthan Radhakrishna, Aman Tiwari, Aanjaneya Shukla, Masoud Hashemi, Rishabh Maheshwary, Shiva Krishna Reddy Malay, Jash Mehta, Pulkit Pattnaik, Saloni Mittal, Khalil Slimi, Kelechi Ogueji, Akintunde Oladipo, Soham Parikh, Oluwanifemi Bamgbose, Toby Liang, Ahmed Masry, Khyati Mahajan, Sai Rajeswar Mudumba, Vikas Yadav, Sathwik Tejaswi Madhusudhan, Torsten Scholak, Sagar Davasam, Srinivas Sunkara, Nicholas Chapados

Open source

Type: other
Venue: arXiv / ServiceNow

Summary

Starts from Pixtral-12B, depth-upscales the decoder 40→48 layers, then two-stage multimodal CPT (foundational text/vision then synthetic visual reasoning) and high-signal text SFT with reasoning traces. No RL or preference optimization. Artificial Analysis Intelligence Index 52 (matches DeepSeek-R1-0528); AIME 2025 87.5, IFBench 61.7, τ²-Bench Telecom 68.4. Ten image benchmarks average within ~5 points of Gemini-2.5-Flash / Claude Sonnet 3.7. Weights MIT at ServiceNow-AI/Apriel-1.5-15b-Thinker.

Keywords

apriel · mid-training · pixtral · multimodal · servicenow · compact-llm · sft · no-rl

Topics

multimodal reasoning, mid-training, compact LLMs

Research notes

Primary: arxiv abs (cs.AI). Paper states model/recipes/evals released under MIT; HF cardData.license mit. ServiceNow SLAM lab. Weights https://huggingface.co/ServiceNow-AI/Apriel-1.5-15b-Thinker (473 likes / 222 downloads at check; usedStorage ~29.7 GB). HF paper page 125 upvotes; linked models not copied into hf_* fields. Discord posted PDF. Training mix is not a standalone public corpus on abs, so no datasets_local row. License field left blank per catalog convention.