What it is
An open 7B foundation model for math and agentic search, trained from scratch with a highly efficient 256K-context recipe and released with code, data, and checkpoints. Instead of scaling parameters, it makes a compact 7B model competitive via deliberate reasoning plus tool use, enabled by an efficient open recipe: hybrid sliding-window/global attention, FP8+Muon pretraining, and MDP-style mid-training on interaction traces.
Key results
- 1st average rank across 14 reasoning benchmarks among 7B-8B models
- 97.1% on MATH-500 and 75.0% on AIME 2026
- ~4.2x 16K pretraining time-to-loss vs BF16/AdamW baseline
Shows that a fully open 7B model can stay competitive on hard math and agentic search by combining efficient long-context training with explicit tool-oriented supervision.