SwarmDo-A1 (release candidate)

SwarmDo-A1 is an open, self-hostable, execution-verified, multimodal coding agent — a LoRA adapter on Qwen/Qwen3.6-27B that fixes real software bugs and writes code from images, and verifies its own patches by running the project's tests. Built in Australia; measured, not marketing — the negatives are published too.

Public model hub & full results: https://github.com/SwarmDo/models

Keywords: open-source coding agent · agentic software engineering · SWE-bench-style patch generation · execution-verified best-of-N selection · image-to-code / visual coding · multimodal vision · Qwen3.6-27B LoRA / PEFT adapter · self-hostable on vLLM · Apache-2.0.

SwarmDo-A1 is a coding system, not just a checkpoint: a strong re-based open base + a hardened agentic harness that verifies patches by running the project's tests, plus a differentiated exec-grounded visual-coding capability (turning a screenshot or mockup into working code and checking it by rendering and comparing). This repository holds the fine-tuned LoRA adapter (the "A1" delta); the base weights are the open Qwen/Qwen3.6-27B. The adapter reaches the architecture's gated-DeltaNet sequence-mixing layers (in_proj_*, out_proj), not just the standard attention projections.

What's decision-grade (paired McNemar, held-out, execution-verified)

Result	Number	Significance
Re-basing to Qwen3.6-27B vs the prior 9B base	≈ +25 pp	p = 0.013 (0.004 stricter)
Base + execution-verified selection vs base alone	36 vs 24 solved (+50%)	p = 0.0005
Visual coding — does the image drive the output?	image-on 15 vs image-off 0	p = 6e-05

These are relative, held-out, execution-verified comparisons on our own task sets — not an absolute public-leaderboard pass rate. We deliberately do not post a single "SWE-bench Verified score" we cannot stand behind under paired testing.

Honest limitations (published, not buried)

• This adapter, as a single model, is ~neutral vs the base. Single-model fine-tuning trades a small quality gain against a robustness cost (a "churn wall"). A1's value is the system — the base plus execution-grounded best-of-N selection, where the win above (p = 0.0005) lives.
• Specialized for Python-centric agentic SWE. Like all coding agents it can produce plausible-but-wrong patches — always run the project's tests against its output. Not a general chat assistant.

Usage

Serve the base with the adapter loaded (vLLM ≥ 0.19 — the base's Qwen3_5 architecture supports adapters on its tower modules):

vllm serve Qwen/Qwen3.6-27B \
  --enable-lora --lora-modules swarmdo-a1=SwarmDo/SwarmDo-A1 --max-lora-rank 32 \
  --tool-call-parser qwen3_xml --reasoning-parser qwen3 --enforce-eager

Then request "model": "swarmdo-a1" on the OpenAI-compatible endpoint (or "Qwen/Qwen3.6-27B" for the raw base). Or load with PEFT:

from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel
base = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3.6-27B", torch_dtype="bfloat16", device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "SwarmDo/SwarmDo-A1")
processor = AutoProcessor.from_pretrained("Qwen/Qwen3.6-27B", trust_remote_code=True)

Serving notes: use --tool-call-parser qwen3_xml (not hermes — it silently drops tool calls); serve in eager mode if you hit hybrid-attention graph-capture issues; the linear-attention base keeps a fixed-size recurrent state, so long contexts cost prefill tokens, not growing KV cache.

License

Apache-2.0, consistent with the base model family.

Citation

@software{swarmdo_a1_2026,
  title  = {SwarmDo-A1: An open, execution-verified multimodal coding system},
  author = {SwarmDo},
  year   = {2026},
  url    = {https://github.com/SwarmDo/models}
}