Inception Labs launched Mercury 3 on October 19, 2026, introducing a diffusion-based AI model now available on Venice.ai. The model generates and refines all tokens in a response simultaneously rather than sequentially, achieving speeds exceeding 1,000 tokens per second on NVIDIA GPUs. The company developed Mercury 3 to address what it identifies as a speed bottleneck in traditional chatbot architectures that generate text one word at a time. This release represents the fourth iteration in Inception Labs' Mercury series, which has focused on advancing diffusion-based language model architectures throughout 2025 and 2026.
Mercury 3 Technical Architecture and Capabilities
Mercury 3 uses a diffusion architecture that drafts an entire response simultaneously, then refines all tokens in parallel passes. Inception Labs reports the model reaches 1,068 tokens per second when running on NVIDIA GPUs.
The model ships with a 1-million-token context window, allowing it to process large amounts of text in working memory during a single task. Mercury 3 supports tool calling, structured outputs, and tunable reasoning, enabling it to trigger external software, return machine-readable formatted answers, and adjust its computational effort.
Inception Labs states Mercury 3 is more than 40 times larger in parameters than Mercury 1 while maintaining similar speed and efficiency metrics.
Benchmark Performance and Speed Metrics
Inception Labs 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. The company states the model shows particular strength in instruction following and agent-based coding while running at 4 to 8 times the speed of those competitors.
These benchmark results are internal and have not undergone independent validation.
An early partnering agent completed 50% of tasks on Mind2Web, a benchmark designed around navigating websites to finish multi-step jobs, according to Inception Labs.
Pricing Structure and API Access
Mercury 3 pricing is set at $0.20 per million input tokens and $0.70 per million output tokens at list rates. A 50% launch discount reduces these rates to $0.10 per million input tokens and $0.35 per million output tokens.
The model is available through API-only access with compatibility for OpenAI's systems. Developers can access Mercury 3 using the model ID "mercury-3". The OpenAI compatibility allows teams using OpenAI's tooling to integrate Mercury 3 without rebuilding their existing infrastructure.
Mercury 3 is offered through Venice.ai under a pay-per-token model. The integration became available shortly after the October 19 launch.
FAQ
What speed does Mercury 3 achieve on NVIDIA GPUs?
Inception Labs reports Mercury 3 reaches 1,068 tokens per second when running on NVIDIA GPUs, using a diffusion architecture that generates and refines all tokens simultaneously rather than sequentially.
What are the pricing rates for Mercury 3?
List rates are $0.20 per million input tokens and $0.70 per million output tokens. A 50% launch discount reduces these to $0.10 per million input tokens and $0.35 per million output tokens.
How does Mercury 3 compare to GPT-6 Luna and Claude Haiku 5.5?
Inception Labs 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 internal benchmarks while running at 4 to 8 times their speed, though these results have not been independently validated.
