CRESCENT

Models · status

Every route, named and checked.

Status shows only whether a server credential is present — never its value. Crescent never swaps a model silently; unavailable routes refuse instead.
Server · …
NVIDIA key · …
Preview AI · …
Receipt signer · …
Self-hosted runtime · …
not configuredvision

Nemotron 3.5 Super VL

nvidia/nemotron-3.5-super-vl-120b-a12b

Provider
NVIDIA hosted API
Best for
Deep reasoning · tool use · vision
Context
262K (hosted endpoint)

The open-weights model card describes up to 1M context; Crescent only advertises the 262K listed for the hosted endpoint.

Model page
not configured

Nemotron 3.5 Lightning

nvidia/nemotron-3.5-lightning-30b-a3b

Provider
NVIDIA hosted API
Best for
Fast text tasks
Context
See NVIDIA model page

Text-only route. Images are rejected instead of being silently re-routed.

Model page
not configuredvision

GPT-6 Astra (Preview AI)

openai/gpt-6-astra

Provider
Lovable AI Gateway
Best for
Opt-in preview route — not Nemotron
Context
Provider-managed

Lets you try Crescent before an NVIDIA key is configured. Every reply and receipt names this model.

runtime required

Bolt 74B-A9B

bolt-74b-a9b (self-hosted)

Provider
Self-hosted runtime
Best for
Local / private deployments
Context
Depends on your runtime

Downloadable model (~24GB). Runtime required: point LOCAL_RUNTIME_URL at your own OpenAI-compatible server. No hosted ID is assumed.

not configured

Hermes Agent runtime

hermes-agent (external runtime)

Provider
External Hermes API server
Best for
Real Hermes harness · skills · memory · tools
Context
Set by the Hermes inference provider

Nous Research's Python agent runs on separate persistent compute (see deploy/hermes). Crescent proxies to it server-to-server with a bearer secret. Hosted Hermes on Nemotron Lightning/Super is cloud inference — not local Bolt.

Model page

Route rules: Fast → Lightning (text only, reasoning off for speed). Deep → Super VL. Auto → Super VL for images or prompts over 1,200 characters, otherwise Lightning. Max output 4,096 tokens per call. Costs are shown as “not priced” — no rates are configured.