HF's 300MB RL LoRA lifts Qwen3.8-27B to 51% on SWE-bench Pro locally
Hugging Face has released its second harness adapter, Qwen3.8-27B-opencode-adapted: a LoRA trained with reinforcement learning on the frozen Qwen3.8-27B, optimized for the OpenCode coding agent. The adapter is only about 300MB, open-sourced under Apache 2.0, and can run locally via llama.cpp. Community members yb2698 and Disastrous-Work-1632, among others, shared the announcement.
Confirmed
- The adapter was released by Aritra of HF and also announced by Lysandre Jik; it is the second harness adapter after the Pi adapter, trained on v2 data with LoRA rank r=32
- Results: pass@1 on the SWE-bench Pro Hard 51 subset rose from 35.3% to 51.0%, a gain of roughly 15.7-16 percentage points
- With thinking mode disabled it still matches xhigh reasoning-level accuracy while saving 6-8x on token consumption (Disastrous-Work-1632 cited a specific figure of 8.1x)
- Fully open source (Apache 2.0), supports local deployment via llama.cpp and more
Why it matters
- A small 300MB LoRA, trained via RL to adapt a general model to a specific coding harness, delivers double-digit benchmark gains and multi-fold token savings—a new paradigm for running strong coding agents locally and cheaply
- "Turning off thinking without losing accuracy" means most use cases can skip lengthy reasoning, further cutting real-world costs
2026-11-18 ~ 2026-11-19 · 5 related posts
- Episode 1: Hugging Face's Harness Adapters: Cheap LoRA+RL Tune Small Models per Agent Framework(2026-11-06, 18 posts)
- Episode 2: Hugging Face Launches Harness Adapters: RL-trained LoRA Boosts Qwen3.8-27B on Pi Coding Agent(2026-11-17, 14 posts)
- Episode 3: HF's 300MB RL LoRA lifts Qwen3.8-27B to 51% on SWE-bench Pro locally(2026-11-18, 5 posts)
Primary sources
- Qwen3.8-27B OpenCode adapter lifts SWE-Bench Pro 35.3%→51% with 6-8x fewer tokens — yb2698 ·
- HF releases 300MB RL-trained LoRA matching Qwen3.8-27B xhigh accuracy with 8.1× fewer tokens — Disastrous-Work-1632 ·
- HF releases RL-trained Qwen3.8 adapter for opencode: +16pts on SWE-bench Pro, 8x fewer tokens — LysandreJik ·
- [source] HF releases 300MB RL-trained LoRA matching Qwen3.8-27B xhigh accuracy with 8.1× fewer tokens — Disastrous-Work-1632 · 2026-11-18
- [source] Qwen3.8-27B OpenCode adapter lifts SWE-Bench Pro 35.3%→51% with 6-8x fewer tokens — yb2698 · 2026-11-18
- OpenCode adapter boosts local Qwen3.8-27B: SWE-Bench 35.3%→51%, 6-8x fewer tokens — yb2698 · 2026-11-18
- [source] HF releases RL-trained Qwen3.8 adapter for opencode: +16pts on SWE-bench Pro, 8x fewer tokens — LysandreJik · 2026-11-18
- 300MB RL-trained LoRA lifts local Qwen3.8-27B coding agent 15.7% on SWE-bench Pro — yb2698 · 2026-11-19