Qwen3.8-27B Super Adapter: Shared by Three Coding Agents

Uploaded by harness-hub on 11 16, this is a LoRA adapter for Qwen3.8-27B. It merges three previously released specialized adapters (each trained via reinforcement learning on software engineering and terminal tasks using AsyncGRPO within the pi, Hermes Agent, and OpenCode coding agent frameworks) by concatenating them with equal weights of ½. The merging process itself involves no further training, and the merged rank is 96. The author states that the weight updates of the three adapters are nearly orthogonal, so most of their individual effects are preserved after addition; this column featured the pi-specific version on 11 18, and the novelty of this version is that a single adapter serves all three frameworks.

Original title: harness-hub/Qwen3.8-27B-super-adapter

Uploaded by harness-hub on 11 16, this is a LoRA adapter for Qwen3.8-27B. It merges three previously released specialized adapters (each trained via reinforcement learning on software engineering and terminal tasks using AsyncGRPO within the pi, Hermes Agent, and OpenCode coding agent frameworks) by concatenating them with equal weights of ½. The merging process itself involves no further training, and the merged rank is 96. The author states that the weight updates of the three adapters are nearly orthogonal, so most of their individual effects are preserved after addition; this column featured the pi-specific version on 11 18, and the novelty of this version is that a single adapter serves all three frameworks.

Why it matters No need to switch adapters based on the agent framework; the author claims that with thinking enabled, five out of six comparisons outperformed the original model while one tied, and the number of tool calls was also reduced

Upvotes 27 · Likes 28 · Downloads 103

Reasons

  • The author claims that with thinking enabled, the average pass@1 across the three frameworks on SWE-bench Pro (51 hard problems) and Terminal-Bench 2.1 increased from 55.5% to 63.5%, with 19%–48% fewer tool calls per problem
  • With thinking disabled, all six comparisons exceeded the original model (average 39.5% → 49.0%); however, if using only one framework with thinking disabled, the author recommends using that framework's specialized version
  • Can be used directly with vLLM by attaching the LoRA, or with llama.cpp using the GGUF version; can be merged via PEFT into a standalone BF16 model of approximately 54 GB

Usage

  • vllm serve Qwen/Qwen3.8-27B --enable-lora --max-lora-rank 128 --lora-modules qwen3.8-super=harness-hub/Qwen3.8-27B-super-adapter (the card also includes parameters for tool calling and reasoning parsing), then point pi, Hermes Agent, or OpenCode to the local OpenAI-compatible interface
  • llama.cpp: Use Q8_0 GGUF as the base, and add --lora Qwen3.8-27B-super-adapter-lora-F16.gguf
  • Keep thinking enabled; in pi's configuration, specify "api": "openai-completions", and Hermes Agent requires at least 64,000 tokens of context

Risks

  • The scores are from the author's single-attempt self-tests within each framework, with a 100-step budget and text-only inputs, and are not comparable to the scores on the official Qwen3.8-27B model card; the ½ merging weight was selected based on Terminal-Bench, and the author admits these results are somewhat optimistic
  • The GGUF version has not been re-evaluated, and low-bit quantized bases may deviate from the evaluated performance; this entry has not been run

Open

Hugging Face Trending

Explore the latest AI Radar updates