UW and Meta unveil Context Language Models that edit their own context

Researchers at the University of Washington and Meta Superintelligence Labs published a paper on Context Language Models (arXiv 2609.37725), where a model edits its context file via Bash commands. Qwen3.6-27B scored 59.4% on BrowseComp-Plus versus 53.4% for Codex-style summarisation while using 21.5% fewer FLOPs. The code is released under CC BY-NC 4.0, with no trained weights.

• Qwen3.6-27B hit 59.4% on BrowseComp-Plus vs 53.4% for summarisation, using 21.5% fewer FLOPs
• RL training lifted Qwen3.5-9B from 28.8% to 42.5% with 38.8% fewer FLOPs
• On EdgeBench 12-hour runs scored 44.6 vs 42.3 with 59% fewer FLOPs
• Code and Suffix Cache Reuse released on GitHub under CC BY-NC 4.0, no weights