Mercury 3 launches on Venice, bringing parallel token generation to private AI
Inception Labs' diffusion-based model claims speeds above 1,000 tokens per second and is now accessible through the privacy-focused Venice platform
Most chatbots write the way a nervous student fills in an exam: one word at a time, never looking back. Inception Labs is betting that approach is a bottleneck.
The company’s new model, Mercury 3, is now live on Venice.ai. It uses a diffusion architecture that generates and refines every token in a response at once, pushing output speeds past 1,000 tokens per second.
What Mercury 3 actually does
Inception Labs launched Mercury 3 on October 19, 2026. It is the fourth model in the company’s Mercury series.
The headline figure is speed. Inception says the model hits 1,068 tokens per second when running on NVIDIA GPUs.
Mercury 3 drafts an entire response simultaneously, then sharpens all of it in parallel passes.
Speed is not the only spec. Mercury 3 ships with a 1-million-token context window, meaning it can hold a very large amount of text in working memory during a single task.
It also supports tool calling, structured outputs and tunable reasoning. Put simply, it can trigger external software, return answers in tidy machine-readable formats, and dial its thinking effort up or down.
Benchmarks, pricing and the fine print
Inception claims Mercury 3 outperforms or matches OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 on 8 to 9 out of 12 benchmarks.
According to Inception, the model shows particular strength in instruction following and agent-based coding. It reportedly does this while running at 4 to 8 times the speed of those rivals.
Those benchmarks are internal, though. Independent validation is still pending.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Inception also says Mercury 3 is more than 40 times larger in parameters than Mercury 1. Despite that growth, the company claims it keeps similar speed and efficiency.
There is one early real-world data point. An early partnering agent reportedly completed 50% of tasks on Mind2Web, a benchmark built around navigating websites to finish multi-step jobs.
On pricing, list rates are set at $0.20 per million input tokens and $0.70 per million output tokens. A 50% launch discount cuts that to $0.10 and $0.35 respectively.
Access is API-only, with compatibility for OpenAI’s systems. Developers can call it using the model ID “mercury-3”.
OpenAI compatibility means teams already built around OpenAI’s tooling can, in theory, swap in Mercury 3 without rewriting their entire stack.
Why Venice is part of the story
Mercury 3 is being offered through privacy-focused platforms, including Venice.ai, under a pay-per-token model.
The integration landed shortly after launch.
How we got here
The Mercury line has been iterating quickly. Prior versions arrived across 2025 and 2026, with each release pushing the diffusion concept further.
What this means
Inception is targeting a specific slice of the market: low-latency work. That includes AI agents, real-time search, and voice and chat interactions.
For developers, the pitch is straightforward. Comparable quality to small frontier models, several times faster, at launch prices of $0.10 and $0.35 per million tokens.
The risk is in the verification. Self-reported benchmarks have a habit of looking better in the press release than in production, and the 8-to-9-out-of-12 claim has not yet been independently confirmed.
The discount is also temporary. Teams that build around the $0.10 and $0.35 rates should budget for the list prices of $0.20 and $0.70 once the promotion ends.
Things to watch next: independent benchmark results, how Mercury 3 handles long tasks that actually use its 1-million-token window, and whether the agent performance seen on Mind2Web holds up beyond an early partner.