Empero Says Quark Learned Tool Use from Reinforcement Alone, Solving 65% of Terminal Tasks
Summary
Empero releases Quark, a 165M-parameter model with 128K context, trained end to end for about $4,200 of compute, roughly 1,300 H100-hours, with four Apache-2.0 checkpoints on Hugging Face. Empero says one version learned tool use from reinforcement learning alone, raising solved terminal tasks from 3% to 65%.
Key Points
- Only four global layers keep a full key-value cache, so at 128K tokens the cache takes about 0.5 GiB in bf16, Empero reports.
- Empero says the GRPO reinforcement learning run takes under 1,300 steps and about 32 hours, with roughly 95% of rewards drawn from code.
- Pretraining runs 500B tokens on a single 8xH100 node, and Empero reports the node goes down seven times with no spike in loss.