SwarmDo Models — open, execution-verified coding agents

Released models

Two checkpoints, one 27B base.

Both are LoRA adapters on the Apache-2.0 Qwen/Qwen3.6-27B base — open, self-hostable, and sized for a single 80GB GPU. A1 is the coding tier; A2 folds a second, independently trained visual-coding skill into the same model.

SwarmDo-A1

release candidate

the coding tier

A strong open base rebuilt for repository-level bug fixing, where candidate patches are selected by running the project’s own tests — never by an LLM judge. Text + vision.

baseQwen3.6-27B
size27B
modalitytext + vision
local runGGUF · Ollama / LM Studio

full page → 🤗 adapter GGUF

SwarmDo-A2

churn-gate cleared

code and vision, one checkpoint

Takes A1’s coding delta and fuses it — on the same 27B base — with a separately trained visual-coding delta, into a single rank-48 adapter. Not two adapters you swap between.

baseQwen3.6-27B
adapterrank-48 LoRA
modalitytext + vision (both load-bearing)
servingmerged, via vLLM

model card → 🤗 adapter merged

Which one? Want a local, code-only model on your own machine — A1, via GGUF. Want code and visual coding from one served endpoint — A2, merged under vLLM. A2 has no GGUF on purpose: quantising to a text GGUF drops the vision tower and makes the visual half inert.

How the gains are earned

Verified by running things, not by vibes.

Every headline number on this site comes from a paired, held-out comparison with a significance test attached. Where a result was null, we publish the null.

+50%

A1 · execution-verified selection

Base plus test-passing patch selection solved half again as many issues as the base alone — 36 vs 24, regression-free.

McNemar p = 0.0005

+25pp

A1 · the base matters most

The single biggest measured win was re-basing onto a stronger open model. Capability dominates cleverness.

McNemar p = 0.013

p = 0.63

A2 · the churn gate

Combining two skills degraded neither. Visual moat vs the pure visual model: Δ −0.006, 95% CI [−0.016, +0.004] — not worse. Coding held too: 18 vs 19 solved, p = 1.0.

neither half eroded

Render-verified visual coding. Show the model a chart; it writes code to reproduce it; we render that code and compare it to the target image. It’s an objective reward you can’t bluff — and the model self-generates its own render-verified training data from it, teacher-free.

Run it yourself

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

No account, no key, no waitlist. If you’d rather not own the hardware, SwarmDoGPU is the hosted path — in development.

# A1 — code-only, local, no account needed
ollama run hf.co/SwarmDo/SwarmDo-A1-GGUF:Q4_K_M
# Q4_K_M ≈ 16.5 GB · Q8_0 ≈ 28.6 GB · LM Studio auto-indexes it

# A2 — code + vision, served merged (vision tower intact)
vllm serve SwarmDo/SwarmDo-A2-merged --served-model-name swarmdo-a2 \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml \
  --reasoning-parser qwen3 --enforce-eager --trust-remote-code

Serving note. Use --tool-call-parser qwen3_xml with --enable-auto-tool-choice — the hermes parser silently drops tool calls, and without auto-tool-choice they 400. Serve once per process; eager mode avoids hybrid-attention graph-capture issues.

What we don’t claim

The negatives, published.

Fine-tuning alone did nothing. Single-model fine-tuning on its own produced no significant gain — a “churn wall”. That is precisely why A1 is a system with a verification loop, not a magic checkpoint.
A2’s coding gate is small. It rules out gross regression; it is underpowered for subtle ones. We say so rather than rounding it into a win.
No leaderboard chasing. Numbers here are paired and held-out with contamination controls. We don’t quote a benchmark we didn’t run.

Full results & reproduce guides The A1 deep-dive Back to the plugin