RC RANDOM CHAOS

Inkling 2 Small Model Adds Vision to Autonomous Work

· via Hacker News

Original source

Inkling 2 Small

Hacker News →

A new Mixture-of-Experts model called Inkling 2 Small has been released with 594 billion total parameters and 17 billion active parameters. The model scored 31 on the Artificial Analysis Intelligence Index with improvements across agentic coding, vision, knowledge work, and reasoning. It includes a 292-million-parameter vision encoder processing 20×20-pixel patches with spatial pooling. The model launches through Tinker with closed weights initially, with full weights scheduled for release under an Apache 2.0 license following external safety testing and red-teaming.

The model demonstrates capabilities in autonomous work spanning coding, research, and knowledge tasks. During training on long-horizon coding tasks, it learned to follow instructions carefully, verify work unprompted, and operate effectively within agent loops. It was tested on real-world coding work and showed strength in autonomous research, including training other models and optimizing for various hardware architectures. In a demonstration, it autonomously trained another model on time series forecasting, achieving a doubled forecasting score after 4.5 hours of work. For knowledge work, it achieved an ELO score of 1527 on a professional benchmark with particular strength in finance and quantitative reasoning.

Safety assessments covered chemical, biological, radiological, and nuclear risks alongside cyber security benchmarks and adversarial testing. The model’s capability scores in CBRN risk remain below existing open-weight frontiers. On cyber security evaluations, it matches frontier performance on two benchmarks measuring vulnerability detection and patching capabilities. Alignment improvements involved filtering training data containing problematic behavior and adding rewards that penalize misbehavior throughout reinforcement learning.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.