Mercury 3 diffusion LLM beats GPT-6 Luna in speed and reasoning

A new speed champion

Inception Labs recently released Mercury 3, a new large language model designed to challenge the dominance of mainstream models like GPT-6 Luna and Claude Haiku 5.5. According to a performance report from eu.36kr.com, Mercury 3 achieves an output speed of 1068 tokens per second on NVIDIA GPUs. This is approximately eight times faster than the reported speed of GPT-6 Luna.

While previous iterations of diffusion models were often criticized for lacking the depth of standard autoregressive models, Mercury 3 aims to bridge that gap. The model is over 40 times larger than the original Mercury 1, yet it manages to maintain its signature high-speed generation. This suggests that the diffusion approach can scale effectively without sacrificing the velocity that makes it unique.

The pricing for Mercury 3 also positions it as a high-efficiency option for developers. During its launch period, the model is offered at $0.10 per million input tokens and $0.35 per million output tokens. These costs are significantly lower than many of its competitors, making it an attractive choice for high-volume applications.

Outperforming the giants

Mercury 3 is not just faster; it is also proving to be more capable in professional and technical domains. On the GPQA Diamond benchmark, which tests graduate-level scientific knowledge and reasoning, Mercury 3 scored 89.23. This outperformed GPT-6 Luna, which scored 86.36, and Claude Haiku 5.5, which scored 84.51.

The model shows even more significant gains in practical software engineering tasks. On the SWE-Bench Atlas benchmark, Mercury 3 achieved a score of 51.08. In comparison, GPT-6 Luna scored 33.33 and Claude Haiku 5.5 scored 32.26. These results indicate that the diffusion route can handle complex instructions and real-world coding problems effectively.

In addition to these reasoning capabilities, the model supports a large context window of 1 million tokens. This is a substantial increase from the 260,000 tokens supported by the previous Mercury 2.5 version. By combining a massive context window with high-speed output, the model is designed to handle long-form data and complex multi-step interactions.

How diffusion works

To understand why Mercury 3 is so fast, it helps to look at how it differs from standard autoregressive models. Models like GPT and Claude predict the next token one at a time in a left-to-right sequence. This creates a serial chain where the model must finish one word before it can begin the next, which limits how much the GPU can do at once.

Diffusion language models take a different approach by predicting the content of multiple positions simultaneously. Instead of writing word-by-word, the model drafts the entire paragraph first and then continuously adjusts it through multiple rounds of iterations. This parallel processing allows the model to update a batch of tokens in a single forward calculation, significantly reducing the number of serial steps required.

The challenge with this method is balancing speed with coherence. If a model performs too many parallel operations without enough refinement rounds, the resulting text can become nonsensical or contradictory. Inception Labs addressed this by using Transformers to make parallel predictions from coarse to fine. This allows Mercury 3 to maintain high quality while still benefiting from the speed of parallel generation.

Impact on AI agents

The release of Mercury 3 has significant implications for the development of AI agents. In agentic workflows, such as a browser agent searching for supplier quotes, the total time to complete a task is the sum of many small delays. Because agents must make dozens of sequential decisions, a faster response time at each individual step results in a much faster overall completion time.

Real-world tests show that Mercury 3 is already being used to improve these experiences. The company TinyFish integrated Mercury 3 into its web proxy system and found it completed 50% of navigation tasks on real websites. This success rate was significantly higher than that of GPT-6.1 Sol, while also consuming fewer tokens and costing less per operation.

Another example comes from Octen AI, which uses the model for real-time search in smart glasses. By switching to Mercury 3, they dropped the median model generation time to 0.9 seconds. This allows for an end-to-end response time of about 2.2 seconds, which makes the interaction feel like a natural conversation rather than a slow request. This highlights a new division of labor where fast models like Mercury 3 act as the 'executors' for complex systems.

The future of scaling

The success of Mercury 3 validates the scaling potential of diffusion language models. It proves that as these models grow in size and complexity, they do not have to lose the speed advantages inherent in their architecture. This opens the door for a new type of model architecture that can handle both high-level reasoning and rapid, low-cost execution.

Tri Dao, the author of FlashAttention, noted that the goal is often to combine a smart 'coordinator' model with a fast 'executor' model. Mercury 3 is positioning itself as a premier executor. It is designed to be smart enough to follow instructions and retrieve information, but fast enough to perform the repetitive, high-frequency actions required by modern AI applications.

Moving forward, the competition between diffusion models and autoregressive models will likely center on finding the right balance of parallelism. As developers look for ways to make AI agents more economical and responsive, models that can execute multi-round operations in seconds will become increasingly valuable. Mercury 3 represents a significant step toward making these high-speed, high-capability agents a reality.