Cumulus
Labs
The end-to-end platform for GPU inference.
One stack, three layers: Paladin orchestrates the fleet, Ion serves at hardware speed, and Talos puts it to work across your enterprise.
Founded by alumni from
Georgia Tech, NASA, Palantir, Space Force, Blackstone, UW Madison
Unified GPU orchestration
Paladin
The control plane for your GPU fleet. Paladin deploys models, partitions GPUs into right-sized slices, and reports occupancy from the silicon itself — observed truth, never inferred.
- Observed-truth occupancy. Fleet state is polled from the GPUs themselves, per device — not inferred from what was requested. The board never lies.
- Partitioning. Slice a GPU into 10–80GB partitions sized to the workload, and hand each slice to a different team or tenant.
- Reconciled deployments. Every model runs a pending → pulling → starting → healthy state machine; endpoints that go dark stop receiving traffic and recover on the next live probe.
› reconciler: qwen2.5-32b pending — awaiting placement
Serverless GPU inference
Ion
Cumulus's serving network for open frontier models. Checkpoint-restored cold starts bring large models up in seconds, behind one OpenAI-compatible endpoint that scales with demand.
- Cold starts in seconds. GPU checkpoint-and-restore loads models 3.4× faster than standard loaders today, with sub-2-second 70B-class starts as the engineering target.
- Open frontier models. Kimi, GLM, Qwen3, MiniMax, DeepSeek — served on Cumulus GPUs behind one OpenAI-compatible API.
- Demand-based capacity. A priority-queue scheduler adds GPUs when contention rises, so bursty traffic never pays for a standing fleet.
Measured, same hardware. Sub-2s cold starts for 70B-class models is the target.
Enterprise agents platform
Talos
Agents that operate your platform with human control built in. Every write an agent wants to make becomes a proposal — nothing mutates until a person applies it.
- Proposal-gated actions. Agent writes are intercepted into proposals; high-risk changes — deploys, routing, model swaps — require explicit approval before they run.
- Working integrations. Slack threads triaged into Linear issues, outbound voice calls over LiveKit/SIP, and forty-plus tools over the platform itself.
- Audit ledger. Every proposal, decision, and application is recorded with its actor, and risky changes carry rollback state.
One platform. GPU to agent.
Paladin
Unified GPU orchestration
The control plane for your GPU fleet. Paladin deploys models, partitions GPUs into right-sized slices, and reports occupancy from the silicon itself — observed truth, never inferred.
Learn moreUse Cases
How teams run Cumulus in production. Full write-ups coming soon.