Mercury 3 hits 1,068 tokens/sec, beating GPT-6 and Claude on 8 of 12 benchmarks
The new model uses parallel generation to outpace rivals in speed and coding benchmarks, though it trails on some retrieval tasks.
Key points
- Mercury 3 achieves 1,068 tokens per second on NVIDIA GPUs with a 1 million token context window.
- It leads GPT-6 Luna and Claude Haiku 5.5 on eight of twelve benchmarks according to Inception data.
- The model is available via API only, with no open weights or quantized builds released.
Inception has released Mercury 3, a diffusion-based language model that generates text in parallel rather than sequentially, achieving speeds of 1,068 tokens per second on NVIDIA hardware. According to MarkTechPost, this release marks the fourth iteration of the Mercury series, following versions 1, 2, and 2.5. The company positions Mercury 3 as a speed-first option for developers building agents, voice bots, and search tools. It features a context window of one million tokens, allowing it to process large amounts of information in a single pass.
The risk
The primary consideration for developers is the trade-off between speed and specific task performance. While Mercury 3 excels in instruction following and agentic coding, it shows weaknesses in long-context retrieval and certain complex coding tasks. According to Inception’s self-reported benchmark table, Mercury 3 leads GPT-6 Luna and Claude Haiku 5.5 on eight of twelve tests. However, Haiku 5.5 outperforms Mercury 3 on DeepSWE v1 and MRCR v2, while GPT-6 Luna edges ahead on AA-Omniscience-Public and AA-LCR v1.1. Teams must evaluate whether the speed advantage justifies potential accuracy drops in specialized workflows.
View the original post on X
Who is exposed
This release targets software developers and engineering teams building automated systems. Inception states that Mercury 3 is more than forty times larger in total parameters than Mercury 1, which launched in 2025. Despite this massive scale increase, the model maintains high speed. The performance gains are most notable in SWE-Bench Atlas, where Mercury 3 scores 51.08, significantly ahead of both rivals. However, the model is not open source. It is available only through an API on Inception-served NVIDIA GPUs, with the specific GPU model undisclosed.
Cost and compliance
The source material does not disclose pricing, licensing fees, or specific compliance certifications for Mercury 3. Inception has not released open weights or quantized builds, meaning organizations cannot self-host the model on their own infrastructure. All performance data cited comes from Inception’s own benchmark table, with every model run at medium settings. The parameter count for Mercury 3 is not disclosed, only that it exceeds Mercury 1 by a factor of forty.
What should you do?
Developers should test Mercury 3 against their current models using their specific workloads to verify the speed benefit. Teams should review the benchmark gaps in long-context retrieval and complex coding to ensure these weaknesses do not impact their applications. Since the model is API-only, organizations must assess the reliability and latency of Inception’s service endpoints before integration.
The background in plain language
Traditional language models generate text one word at a time, waiting for each step to finish before starting the next. Diffusion models start with a rough draft of the entire answer and refine it in parallel. This approach reduces waiting time, allowing faster output on the same hardware. Teams usually evaluate whether this speed gain holds up as models grow larger and more complex. Check whether your tasks require sequential precision or can benefit from parallel processing.
What to do and how to stay safe: Inception
- Run internal benchmarks comparing Mercury 3 against your current model for your specific use cases.
- Monitor API latency and uptime if you integrate this third-party service into production workflows.
- Review the model’s performance on long-context tasks to identify potential accuracy risks.
- Verify that the API provider’s data handling policies meet your organization’s security requirements.
Step-by-step guide: AI Governance Explained: Building Rules for Autonomous Systems
General security guidance from the Vector Tribune newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How fast is Mercury 3 compared to other models?
Inception reports Mercury 3 runs at 1,068 tokens per second on NVIDIA GPUs, significantly faster than traditional autoregressive models.
Does Mercury 3 outperform GPT-6 Luna and Claude Haiku 5.5?
According to Inception’s benchmarks, Mercury 3 leads on eight of twelve tests, but trails on DeepSWE v1 and MRCR v2.
Is Mercury 3 available for self-hosting?
No, Inception offers Mercury 3 via API only, with no open weights or quantized builds available for local deployment.
