Inception Launches Mercury 3 Diffusion Model

Inception launched Mercury 3 on the 19th local time, describing it as a diffusion large language model (dLLM) that generates 1,068 tokens per second and delivers higher quality than GPT-6 Luna and Claude Haiku 5.5. Based on its own benchmarks, the company called it the strongest and most intelligent dLLM released so far. Inception said the model is more than 40 times larger than the previous generation while maintaining its speed.

Unlike autoregressive models, which generate tokens one after another, Mercury 3 uses diffusion: it drafts an answer, then refines tokens in multiple positions at the same time. Inception said this approach can reduce response latency by cutting sequential processing steps. The model has more than 40 times the parameters of the first model, Mercury 1 Mini, and the two models’ average performance on standard benchmarks improved about threefold. The company did not disclose the parameter counts.

Inception’s evaluations gave Mercury 3 a score of 51.08 on SWE-bench Atlas, ahead of GPT-6 Luna at 33.33 and Haiku 5.5 at 32.26. It scored 76.84 on IFBench, about 15 points above GPT-6 Luna, and 89.23 on GPQA Diamond. Haiku 5.5, however, scored 51.62 on DeepSWE v1, above Mercury 3’s 46.31, and also led on the long-context retrieval benchmark MRCR v2. Inception acknowledged room to improve on some long-document retrieval and complex coding tasks.

The company said Mercury 3 can serve speed-sensitive AI agent tasks such as code review, interactive education, web browsing, real-time search, and customer support. Agents make repeated model requests as they work, so delays at each step can accumulate. Inception said fast responses can speed up web browsing and document analysis and allow agents to try several approaches within a limited time. Early-access partner TinyFish said it completed 50% of 50 real-world web-browsing tasks. Octen AI reported a median generation time of 0.9 seconds and a total response time of 2.2 seconds for smart-glasses search; RevisionDojo said it delivers learning hints and answer reveals in about 1 second.

Mercury 3 is available through Inception’s API platform, OpenRouter, Baseten, and Vercel. It is compatible with the OpenAI API and supports context lengths of up to 1 million tokens. Standard pricing is $0.20 per million input tokens and $0.70 per million output tokens. Launch promotional prices are $0.10 and $0.35, respectively, and cached input tokens receive a 90% discount from the input price. Enterprise customers can access dedicated processing capacity, autoscaling, and data-retention settings. Inception also unveiled Mercury Voice, designed to reduce first-response latency for voice AI, and Mercury Decide, a decision model for large-scale classification.

Inception launched Mercury 3 on the 19th local time, describing it as a diffusion large language model (dLLM) that generates 1,068 tokens per second and delivers higher quality than GPT-6 Luna and Claude Haiku 5.5. Based on its own benchmarks, the company called it the strongest and most intelligent dLLM released so far. Inception said the model is more than 40 times larger than the previous generation while maintaining its speed.

Unlike autoregressive models, which generate tokens one after another, Mercury 3 uses diffusion: it drafts an answer, then refines tokens in multiple positions at the same time. Inception said this approach can reduce response latency by cutting sequential processing steps. The model has more than 40 times the parameters of the first model, Mercury 1 Mini, and the two models’ average performance on standard benchmarks improved about threefold. The company did not disclose the parameter counts.

Inception’s evaluations gave Mercury 3 a score of 51.08 on SWE-bench Atlas, ahead of GPT-6 Luna at 33.33 and Haiku 5.5 at 32.26. It scored 76.84 on IFBench, about 15 points above GPT-6 Luna, and 89.23 on GPQA Diamond. Haiku 5.5, however, scored 51.62 on DeepSWE v1, above Mercury 3’s 46.31, and also led on the long-context retrieval benchmark MRCR v2. Inception acknowledged room to improve on some long-document retrieval and complex coding tasks.

The company said Mercury 3 can serve speed-sensitive AI agent tasks such as code review, interactive education, web browsing, real-time search, and customer support. Agents make repeated model requests as they work, so delays at each step can accumulate. Inception said fast responses can speed up web browsing and document analysis and allow agents to try several approaches within a limited time. Early-access partner TinyFish said it completed 50% of 50 real-world web-browsing tasks. Octen AI reported a median generation time of 0.9 seconds and a total response time of 2.2 seconds for smart-glasses search; RevisionDojo said it delivers learning hints and answer reveals in about 1 second.

Mercury 3 is available through Inception’s API platform, OpenRouter, Baseten, and Vercel. It is compatible with the OpenAI API and supports context lengths of up to 1 million tokens. Standard pricing is $0.20 per million input tokens and $0.70 per million output tokens. Launch promotional prices are $0.10 and $0.35, respectively, and cached input tokens receive a 90% discount from the input price. Enterprise customers can access dedicated processing capacity, autoscaling, and data-retention settings. Inception also unveiled Mercury Voice, designed to reduce first-response latency for voice AI, and Mercury Decide, a decision model for large-scale classification.