Hugging Face ships harness adapters: RL-tuned Qwen3.8-27B gains on Terminal-Bench 2.1
yb2698 · x · 2026-11-21
Hugging Face introduces "harness adapters" — adapt your favorite model to your harness to improve performance and cost. The team has started RL training of Qwen3.8-27B with thinking mode, with encouraging early signals.
Current results (TB-2.1, 68 tasks):
- The step-70 think harness adapter beats the base think mode and slightly beats the non-think Pi adapter + think combo;
- It's more token-efficient than base think and more efficient than the previous Pi adapter;
- Screenshot data: vs base think32k, Checkpoint-70 hits 67.6% (+11/-2 tasks, p=0.022) with -23% tool calls and -21% tokens; vs think32k-pi-mt, 70.6% (+13/-3, p=0.021) with -38% tool calls at only +3% tokens.
The companion article covers generating custom RL tasks via openenv, testing filters with short runs, and AsyncGRPOTrainer for efficient GRPO training.
More from coding & agent
- isometric-cities: coding agents turn real neighborhoods into pixel-art maps on Mac — zsakib_ · 2026-11-21
- Anthropic's MCP Code Mode Cuts Agent Tool Context From 150K Tokens to 2K (98.7% Savings) — blaizedsouza · 2026-11-21
- Grok Bot template turns one idea into a coded animated video with music — nima_owji · 2026-11-21
- Fake font files in 21 GitHub repos hide malware auto-run via VS Code tasks — JeremyCMorgan · 2026-11-21
- Japanese tech lead argues code design matters more, not less, in the AI coding era — 4310sy · 2026-11-21
- Google open-sources stitch-skills: 16 design skills for AI coding tools, 11k GitHub stars — lxfater · 2026-11-21