Artificial intelligence / news brief
IQuest details Q1's 29-trillion-token training pipeline
- Event
- Published
IQuest published a technical report for Q1, its existing 320B-parameter sparse MoE model for command-line agents. The paper describes 29 trillion pre- and mid-training tokens, multi-harness reinforcement learning and results across 30 benchmarks; the scores remain author-reported, and the model card warns that long CLI tasks still require human oversight.
Why it matters
The report adds methods and evaluation detail to an already released open-weight agent model, making its training claims easier to inspect without independently reproducing them.
Sources
- IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems (external site)IQuest Research · primary source ·
- IQuest-Q1 opens 320B/15B MoE, 83.2 on Terminal-Bench 2.1 (external site)AI Weekly · coverage ·