SwarmDo-A2 — merged weights (turnkey, code + vision in one model)
Pre-merged, ready-to-serve weights of SwarmDo-A2. The
combined coding + render-verified visual-coding adapter is already merged into Qwen/Qwen3.6-27B with the
vision tower intact, so you can serve it directly with vLLM — no merge step needed.
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
Then call it as an OpenAI-compatible chat endpoint (model = "swarmdo-a2"); pass an image via the standard
image_url block for visual-coding (image → code).
One model, both skills — fixes real bugs (execution-verified) and writes code from images (render-verified), with measured proof that combining the two degrades neither (the "churn gate").
- Full model card, results & methodology: SwarmDo/SwarmDo-A2
- The adapter (LoRA form, for re-merging with your own settings): same repo
- Public hub: github.com/SwarmDo/models (
a2/)
Notes: serve as these merged weights (native fast path). Do not load A2 as a vLLM LoRA — the adapter reaches the base's gated-DeltaNet modules, which JIT-hang under vLLM's LoRA path. There is no text GGUF for A2 (quantizing to GGUF drops the vision tower). Apache-2.0.