Foundation models for agent workloads
Essential AI develops foundation models, led by CEO Ashish Vaswani. Its Rnj-1 work focuses on coding, tool use and other tasks involved in agents. The company describes a shift toward pretraining in 2025. [1][2]
In December 2025 it released Rnj-1, an 8-billion-parameter model family with base and instruction-tuned versions. Essential AI describes training the models from scratch. Following a similar architecture to another model family does not make these checkpoints fine-tunes of that family's weights. [2]
Public weights with a bounded date
The instruction model card provides downloadable weights under Apache 2.0, with a 32,000-token context window and instructions for local inference. Public checkpoints let developers inspect and adapt the model, rather than depending solely on hosted access. Benchmark scores and training descriptions remain the developer's account. [3][2]
The announcement and model card disagree on the initial release day. This chronology therefore records December 2025 without selecting a day that the sources do not consistently support. [2][3]
Sources
- Essential AI
Essential AI · retrieved Oct 23, 2026
- Announcing Rnj-1: Building Instruments of Intelligence
Essential AI · Dec 5, 2025
- Rnj-1 Instruct Model Card
Essential AI · retrieved Oct 23, 2026