Replicate

Available

Replicate has 14 models tracked on Modelglass across 3 verticals (Image, Video, 3D), most recently added 2025-03.

Replicate has 10 image models (7 current-gen, 3 previous-gen), 3 video models (2 previous-gen, 1 deprecated), and 1 3d model on Modelglass, most recently added 2025-03.

replicate.com ↗

Coverage across verticals

Image 10 models

7 current-gen · 3 previous-gen

Image pricing page ↗

Hosts many open-weight models; most public models are billed per second of GPU time by hardware. Per-model pages also list pricing.

Video 3 models

2 previous-gen · 1 deprecated

Video pricing page ↗

Replicate charges per-second of GPU time consumed, billed at published hardware rates. For video generation models this typically translates to a per-run (per-clip) price that varies with clip duration and resolution. Per-run estimates recorded in this registry are based on typical generation parameters.

3D 1 model

1 ungraded

3D pricing page ↗

Inference aggregator. Hosts open-weights 3D-generation models billed by hardware-second, which Replicate surfaces as an approximate per-run cost on each model page ("this model costs approximately $X to run"). Relevant models: tencent/hunyuan3d-2 (v2.0, ~$0.024/run on L40S, ~25s) and tencent/hunyuan-3d-3.1 (v3.1, up to 8 input views). No subscription; pay-as-you-go against account credit.

See how Replicate's models fit into a cost-aware routing setup

Release cadence

14 models tracked across all verticals. Earliest release 2023-07, most recent 2025-03.

Derived from Modelglass's per-vertical provider registries — image/llm/video/audio's registry/providers/*.yaml records, plus science/coding/agentic/gdpval's benchmark registries, where a provider is identified by the prefix before "/" in each model's model_id rather than a registry record of its own. Where the same company uses a different provider slug or model_id prefix across verticals (e.g. Google DeepMind vs. Google Cloud, or Qwen's "alibaba" model_id prefix in the coding registry vs. its "qwen" provider slug elsewhere), coverage is merged onto a single profile per an explicit, human-reviewed alias list (registry/provider-aliases.yaml) — everything else reflects exact-slug or exact-prefix matches only.