Venice Lists Mercury 3 for API Use With Prompts Visible to Providers

Venice has listed Inception Labs' Mercury 3 for API access, giving developers a new route to the diffusion model while its anonymized mode still exposes prompt content to the model provider.

Inception Labs released Mercury 3 on Oct. 19. Venice identifies the API model as mercury-3 and makes it available through a route compatible with the OpenAI API. The model has been available on Venice since October, with no specific start day. That compatibility lets developers use an existing OpenAI API client format to send requests.

Mercury 3 has a one-million-token context window, the amount of material it can consider in a request. Venice lists a maximum output of 128,000 tokens. The model uses parallel token refinement, revising multiple tokens together, while autoregressive models produce text one token at a time. The design is aimed at reducing waits in workflows that make repeated model calls, such as web agents and search tools.

Standard prices are $0.20 per million input tokens and $0.70 per million output tokens. The launch discount cuts those rates to $0.10 and $0.35, respectively, with no stated end date.

Mercury 3’s published speed is 1,068 tokens per second on NVIDIA GPUs. Token-generation speed alone does not establish how quickly an entire task finishes or whether an answer is accurate.

A task-level evaluation conducted Oct. 19 and 20 recorded a median 0.42 seconds per timed chess move with reasoning off. One move with reasoning on took 105 seconds. The model's analysis had at least one fault in each of the 12 survey tasks.

Venice's anonymized mode hides a user's identity from the model provider, but the provider can still see prompt content. The anonymized label does not mean Mercury 3 runs in a trusted execution environment (TEE) or uses end-to-end encryption (E2EE).

An earlier Mercury 3 code-generation test focused on map checks examined a separate use of the model.