Run agents on managed frontier models today—or rent GPU capacity, deploy your own weights, and route Digio tasks to private endpoints in the same workspace.
Assign a default model per agent or override per task. Usage is metered in Digio Tokens from your plan balance—the same wallet whether the agent calls Sonnet, GPT-4o, or Gemini Flash.
Model list and token economics evolve with provider releases. Your workspace shows live options when you assign a model to an agent; Digio Tokens debit from the same balance as in pricing.
The Coordinator can recommend Sonnet vs Opus vs a cheaper flash model based on task type. Power users set defaults per agent role—research on Sonnet, final review on Opus, bulk tagging on Haiku or Gemini Flash.
Per agent — default model in agent settings; override in To do or chat when needed.
Metered fairly — input, output, and cached tokens map to Digio Token charges (see usage in your wallet).
Skills stay the same — tools and integrations work across models; only latency and cost profile change.
Plan limits — more agents and monthly Digio Tokens on higher tiers; top up anytime on the pricing page.
Need a fine-tune, an air-gapped checkpoint, or predictable inference pricing? Add dedicated GPU capacity to your Digio workspace, install the serving stack you prefer, and point agents at your private endpoint.
Hourly or monthly GPU nodes (A100, H100, L40S class) attached to your tenant—isolated from other customers.
Upload safetensors, GGUF, or pull from your registry; run Llama, Mistral, Qwen, and custom fine-tunes.
vLLM, TGI, Ollama, or container images you maintain—Digio agents call an OpenAI-compatible base URL.
To do, team chat, skills, and collaboration unchanged—only the inference backend is yours.
Send sensitive steps to private GPU and use Claude or GPT for public research in one workflow.
VPC peering, static egress, audit logs, and model allowlists for regulated teams.
Typical setup from zero to agents calling your endpoint:
Choose VRAM, region, and uptime (burst vs always-on). Storage for weights ships with the instance or mounts your bucket.
Start a serving image or SSH in, install CUDA drivers, and load checkpoints. Health checks confirm the model is ready.
Add base URL, API key, and model id in workspace settings. Digio validates latency and token format before going live.
Pick your private model as the default for selected agents; managed Claude/GPT models remain available side by side.
GPU rental is billed separately from Digio plan subscriptions. Contact us for capacity planning, SLAs, and migration from an existing inference cluster.
Choosing managed APIs vs self-hosted inference on Digio.
Your Digio subscription covers infrastructure, agents, and included Digio Tokens. Managed model usage debits that token balance by actual input/output tokens. GPU rental is an add-on for the machines you control.
Yes—each agent can have its own default. Tasks and chats can override for a single run without changing the global default.
Opus is tuned for harder reasoning and longer coherent plans; Sonnet is faster and cheaper for everyday agent loops. Haiku and flash-class models are best for volume subtasks.
Enterprise workspaces can restrict outbound model providers and route all agent traffic to your GPU endpoint. Hybrid mode is the default for most teams.
Offerings depend on region and demand—commonly 24–80 GB VRAM tiers for 7B–70B class models and multi-GPU nodes for larger stacks. We help size VRAM from your parameter count and quantization.
Orchestration (agents, tasks, storage) stays on your plan. Inference on your GPU is billed as GPU time; you may optionally meter token-shaped usage for internal chargeback.
Agents, tasks, and skills—the workspace around the model you choose.
Start on Claude and GPT today, then add dedicated GPU when you are ready to host custom weights—same agents, same tasks, your inference.
