Inception Launches Mercury 3 Diffusion LLM With 1,068 Tokens per Second Speed
The new diffusion-based model claims superior speed and benchmark performance over rivals GPT-6 Luna and Claude Haiku 5.5 in self-reported tests.
Key points
- Mercury 3 generates 1,068 tokens per second on NVIDIA GPUs with a one-million-token context window.
- Inception reports the model leads competitors on eight of twelve benchmarks, including significant gains in agentic coding.
- The model is available via API only, with no open weights or quantized builds disclosed by the vendor.
Inception has released Mercury 3, the latest iteration of its diffusion-based large language model. According to MarkTechPost, the system achieves a generation speed of 1,068 tokens per second on NVIDIA GPUs while supporting a context window of one million tokens. This release marks the fourth major version in the Mercury series, following earlier launches in 2025 and subsequent updates that introduced tunable reasoning and smaller context limits.
The company positions Mercury 3 as a high-speed alternative for developers building agents, voice bots, and search tools. Inception states that the new model is more than forty times larger in total parameters than the original Mercury 1. Despite this massive increase in size, the model maintains speeds above one thousand tokens per second. Average quality scores across standard benchmarks reportedly rose approximately threefold between the initial mini version and this current release.
View the original post on X
Technical Performance and Benchmark Results
Inception’s own benchmark table indicates that Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 on eight of twelve tested metrics. All models were evaluated using medium settings. The most significant advantage appears in agentic coding tasks, where Mercury 3 scored 51.08 on SWE-Bench Atlas. This result places it roughly eighteen points ahead of both rival models, which scored 33.33 and 32.26 respectively.
However, the model shows weaknesses in other areas. Claude Haiku 5.5 surpassed Mercury 3 in DeepSWE v1 with a score of 51.62 compared to 46.31. Haiku also led in MRCR v2 with 78.40 versus 75.50. GPT-6 Luna edged ahead in AA-Omniscience-Public and AA-LCR v1.1 tests. These results suggest that long-context retrieval and longer coding tasks remain challenging for the diffusion approach.
Model Architecture and Availability
Unlike traditional autoregressive models that generate text one token at a time, diffusion LLMs start with a rough draft of the full answer. The system then refines every position in parallel over several passes. This method reduces sequential steps, allowing for faster output on the same hardware. The open question in the industry was whether this parallel approach would maintain quality as models grew larger.
Mercury 3 is currently available through an API only. Inception has not disclosed open weights or quantized builds for the model. The exact NVIDIA GPU configuration used for testing was not specified. The release follows a lineage that includes Mercury 1, Mercury 2, and Mercury 2.5, with the latter adding a 260K context window and tunable reasoning capabilities.
What should you do?
Security and technology professionals should monitor the integration of high-speed models into automated workflows. Rapid generation speeds can increase the volume of output in agentic systems, potentially amplifying errors if not properly validated. Teams using AI for coding or search should verify that the speed benefits do not compromise accuracy in critical long-context tasks.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Inception
- Evaluate current AI agent workflows to ensure that increased generation speeds do not bypass existing validation checks or safety filters.
- Review vendor benchmark claims independently, particularly for models that report superior performance in specific coding or reasoning tasks.
- Monitor API usage patterns for anomalies, as high-throughput models may introduce new vectors for resource exhaustion or prompt injection attacks.
- Verify that your organization’s data handling policies account for the large context windows available in newer models like Mercury 3.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How does Mercury 3 differ from traditional LLMs?
Mercury 3 uses a diffusion approach that refines a rough draft in parallel, rather than generating tokens sequentially one by one.
Is Mercury 3 open source?
No, Inception offers the model via API only and has not released open weights or quantized builds.
Which benchmarks did Mercury 3 perform poorly on?
It scored lower than competitors on DeepSWE v1, MRCR v2, AA-Omniscience-Public, and AA-LCR v1.1 tests.
