News · 2026-10-20

Inception launches Mercury 3, betting parallel text refinement can cut agent waiting time

Inception Labs launched Mercury 3 on October 19, reporting text generation at 1,068 tokens per second through a model that refines multiple tokens in parallel. The hosted release targets interactive applications and agent subtasks where waiting for generated text slows the workflow. Its speed and quality comparisons are company-reported results, rather than independent evidence that it completes every task faster or better.

Key facts

• Inception reports 1,068 generated tokens per second on NVIDIA graphics processors.
• Mercury 3 launched October 19 with a million-token context window.
• Standard pricing is $0.20 per million input tokens and $0.70 per million output tokens; launch pricing halves those rates.
• The primary source is Inception's Mercury 3 announcement.

A conventional text generator resembles a typist moving steadily from the beginning of a sentence to its end. Mercury's approach is closer to drafting a rough page and repeatedly improving several parts of it together. Inception calls its launch “Diffusion at Scale,” emphasizing that the method can produce more than a small, fast demonstration model.

The company's earlier mechanism explainer describes coarse-to-fine refinement: a transformer proposes changes to multiple tokens at once. The process still requires repeated neural-network computations. Parallel refinement does not mean an entire answer appears instantly, and it does not remove the need to judge whether the final answer is correct. Our diffusion language model lesson explains why this is a different generation process rather than a faster setting on an ordinary left-to-right model.

The scaling claim has a precise boundary. Inception says Mercury 3 has more than 40 times Mercury 1's total parameters and reports a threefold improvement on its benchmark-measured intelligence comparison with Mercury 1 mini. It does not disclose the new model's absolute parameter count in the announcement. Nor does that 40-times figure describe customer traffic, production volume or the number of requests the service can sustain.

Inception also says Mercury 3 runs four to eight times faster than named speed-focused competitors while matching or exceeding their quality across a range of benchmarks. That is a useful proposition to test, but the announcement does not provide enough protocol detail to reproduce the headline comparison independently. Output throughput, delay before the first useful response, concurrency and end-to-end task completion are different measurements.

Consider a browser agent that spends several seconds waiting for a page, then emits a short instruction. Generating that instruction much faster may save little overall. An interactive application producing a long explanation could benefit more. A fast model that needs an extra failed attempt can also lose its nominal advantage. The right question is which part of a particular workflow the model actually accelerates, a distinction developed in our end-to-end speed lesson.

The launch discount also belongs in a cost comparison: promotional rates and standard rates are distinct. A workload assessment should record which price was used alongside the model version and test conditions.

The product is available through Inception's interface and, according to its announcement, OpenRouter, Baseten and Vercel. It uses an OpenAI-compatible application interface, which should reduce integration work for applications already built around that request shape. Compatibility still does not guarantee identical prompting behavior, tool performance or error handling, so a switch deserves a representative evaluation.

This is a hosted release, rather than a downloadable checkpoint. In the Hacker News launch discussion, Inception co-founder Volodymyr said weights were not being released at this time. A local disk-size or graphics-memory claim would therefore misdescribe what customers can obtain. The announcement's early partner examples include browser agents, tutoring, search and support, but do not establish large-scale request volume or fleet utilization.

The same discussion supplies a valuable qualification to the marketing. The co-founder said Mercury 3 should not yet be expected to perform at Opus level as a primary agent, while describing it as useful for coding subtasks. Practitioners asked about response latency and performance at larger batch sizes; detailed batch curves were not supplied in the thread. The disagreement is about how speed translates into useful capability, rather than whether the interface exists.

Mercury 3's significance is that an alternative text-generation method is being offered as a practical commercial service. A developer can test whether parallel refinement makes a time-sensitive workflow feel different. The caveat is equally concrete: the public evidence supports an available product and vendor benchmarks, while independent comparisons of successful tasks, realistic concurrency and operating cost remain the next test.

Primary source, verified: read the paper →

Ground Truth. (2026, October 20). Inception launches Mercury 3, betting parallel text refinement can cut agent waiting time. Ground Truth. https://groundtruth.day/news/mercury-3-diffusion-language-model-api-launch.html

@misc{groundtruth:mercury-3-diffusion-language-model-api-launch,
  title  = {Inception launches Mercury 3, betting parallel text refinement can cut agent waiting time},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/mercury-3-diffusion-language-model-api-launch.html}
}

Topics: inception · diffusion · models · agents · inference