SubQ

SubQ (Subquadratic) serves two products: SubQ Snap, a multimodal decision model (384K context, text and image input, typed answers with probabilities), and subq-preview, an early-access long-context chat model. They configure differently in UGENT: Snap belongs under Decision Models; the chat model is an openai-compatible instance with one SubQ-specific rule.

Decision Model (SubQ Snap)

[decisions.instances.subq]
provider = "subq"
api_key = "$SUBQ_API_KEY"

The preset fills https://api.subq.ai, /v1/systemone, and subq-snap-1.1. Usage, limits, and the decide tool are covered in Decision Models. Pricing: $0.04 per million input tokens, output free, 100M free tokens for new accounts.

Chat Model (subq-preview)

SubQ's chat endpoint speaks the OpenAI wire format with one documented exception: requests must send max_tokens, and max_completion_tokens is refused. Set legacy_max_tokens = true and UGENT serializes the budget under the field the endpoint takes:

[llm.instances.subq-chat]
type = "openai-compatible"
api_key = "$SUBQ_API_KEY"
base_url = "https://api.subq.ai/v1"
default_model = "subq-preview"
legacy_max_tokens = true
max_tokens = 64000

Notes:

• Pricing and the context window are unpublished (early access). UGENT's catalog carries the model with placeholder figures and no cost estimates until SubQ publishes.
• max_tokens is a spend cap from which reasoning tokens are drawn; a turn that exhausts it can return an empty answer under finish_reason: "length". Only reasoning_effort: none guarantees no reasoning; UGENT's [reasoning] block maps to that field (none through max).
• Retention: prompts and completions are deleted within 60 days; organization-level zero retention is available on request.
• A type = "subq" instance is rejected with an error pointing at this page — the subq provider name exists so the pricing catalog maps.