SwarmDoGPU

Where it fits

Three layers. This one is the floor they stand on.

SwarmDo has always been two halves — the orchestration plugin and the open models. Both of them ultimately need the same thing: a GPU to run on. SwarmDoGPU is that layer, offered as a hosted service.

Layer 1 · orchestration

swarmdo

The plugin and MCP server. Coordinates agents, memory, and consensus inside Claude Code. Runs on your laptop.

shipped

Layer 2 · models

SwarmDo-A1 / A2

Open, self-hostable coding models. Apache-2.0 base, fits a single 80GB GPU — if you have one.

released

Layer 3 · compute

Managed GPU workloads and infrastructure for the ecosystem above — so “self-hostable” doesn’t have to mean “self-hosted”.

in development

The gap it closes. SwarmDo-A1 and A2 are genuinely open — Apache-2.0 weights, no API lock-in. But “fits on one 80GB GPU” still assumes you have an 80GB GPU. A2 additionally has to be served merged under vLLM, with specific parser and eager-mode flags, because its adapter reaches the base’s gated-DeltaNet layers. That is a real operational cliff, and it is the cliff this service exists to remove.

What it’s being built for

The workloads the ecosystem already generates.

These are the jobs SwarmDo itself needs compute for today — documented in the model cards and the plugin’s own runtime. They are the design targets, not a shipped feature list.

◉Serving A1 & A2

An OpenAI-compatible endpoint for the SwarmDo models, with A2’s merged-weights and tool-parser requirements handled for you rather than rediscovered from a README.

✓Execution-verified selection

The models earn their gains by running candidate patches against a project’s own tests. That sampling-and-testing loop is inherently parallel, and inherently hungry for compute.

🖼Render-verified visual coding

A2’s visual half is scored by rendering generated code and comparing it to the target image — another verify loop that wants throughput more than it wants a single fast card.

⛁Embeddings & vector memory

The plugin’s persistent memory runs ONNX embeddings and HNSW indexing locally today. Larger swarms and shared team memory push that off the laptop.

Honest status

What we don’t claim yet.

SwarmDo publishes null results and unfinished work rather than burying them. The same rule applies to this page.

No product is live. The SwarmDoGPU repository currently holds project scaffolding and integration config. Application code, usage instructions, and deployment docs are still to come.
No pricing is announced. There are no published tiers, rates, or limits — anything you read elsewhere claiming otherwise isn’t from us.
No benchmarks are claimed. We haven’t measured this service, so we won’t quote numbers for it. The numbers on this site belong to the models, and every one of them is reproducible.
No hardware is promised. Specific GPU types, regions, and availability will be stated when they’re real.

In the meantime, both other layers work today and are fully open: the plugin is npx swarmdo init, and the models are downloadable weights you can serve on your own hardware with no account at all.

Early access

Tell us what you’d run on it.

The build is shaped by real workloads. If you want to serve A1 or A2 without buying a card, run execution-verified selection at scale, or push swarm memory off your laptop — say so, and you’ll be in the first group we talk to.

Email for early access See the models it serves Back to the plugin

maintainers@swarmdo.com · no list, no drip campaign — a human replies.