SwarmDo-A1 — open, execution-verified multimodal coding model

What's decision-grade

Measured, paired, and significant — or not claimed.

Every number is a paired statistical test (McNemar exact p) on a contamination-controlled held-out set, reconciled for infrastructure losses. No cherry-picked demos.

+25pp

Re-base win

Rebuilding on a stronger open base (Qwen3.6-27B) beat the prior base by a wide margin. Capability dominates cleverness.

McNemar p = 0.013

36 / 24

Verify by running tests

Sampling candidate patches and keeping the ones that pass — the harness's real value-add. +50% more issues solved.

McNemar p = 0.0005

p=6e-05

Visual coding is real

Given a chart image, the model writes code to reproduce it — scored by rendering & comparing. The image genuinely drives the output.

image-on 15 vs off 0

The differentiator

An exec-grounded visual-coding moat that fuels itself.

Show it a picture; it writes the code; we render that code and compare to the target — an objective reward you can't bluff. It uses two assets few open models have together: a vision tower and an execution verifier.

Greenlit. The image is measurably load-bearing — image-on solves charts image-off cannot (p = 6e-05), render-reward +0.088 (CI excludes zero).

It self-fuels. From 150 charts the base produced render-verified-good code on 63% — with no external teacher. That's the raw material for a self-improving loop.

01Generate — best-of-N image→code from the vision-active model.
02Render — run each candidate in a sandbox; produce the image.
03Score — compare to the target, 0–1. Objective, un-bluffable.
04Keep & train — feed the high-scoring reproductions back. Repeat.

Run it yourself

Open weights, one GPU, no API lock-in.

A1 ships as a LoRA adapter on the open Qwen/Qwen3.6-27B base. Serve it through vLLM as an OpenAI-compatible endpoint.

# serve the base with the A1 adapter (single 80GB GPU, vLLM ≥ 0.19)
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 call model "swarmdo-a1" on the OpenAI-compatible endpoint

On-device paths exist too: MLX (4-bit, ~16GB, vision kept), GGUF (llama.cpp / Ollama), and GPTQ-Int4 vLLM on a single 24GB GPU. Apache-2.0.

Honest by design

What we don't claim.

No magic single-model gain. Fine-tuning one checkpoint alone was ~neutral — a "churn wall" that trades quality for robustness. So A1 is a system (strong base + execution-verified selection), not a magic checkpoint.
SWE-bench absolutes are inflated (it's largely in pretraining data) — only the paired deltas above are trustworthy.
The visual flywheel builds & trains but doesn't yet compound. One small round is roughly neutral on unseen charts; making it compound is the next honest step, not a claimed win.

Open, execution-verified, and Australian. 🇦🇺

Ships alongside the SwarmDo plugin — agent-swarm orchestration for Claude Code — as one open agentic-coding stack.

Download SwarmDo-A1 Reproduce the results