Inception Releases Mercury 3 Diffusion LLM: 1,068 Tokens/Second, Million-Token Context

Diffusion Models Enter the LLM Inference Race

Inception has officially released Mercury 3, a large language model built on a diffusion architecture. Official figures put its inference speed at 1,068 tokens/second, with support for a 1 million token context window. Both numbers are quite aggressive in the context of today's mainstream autoregressive models—most commercial models still run in the range of tens to hundreds of tokens per second, and million-token contexts typically appear only in a handful of flagship models.

Mercury 3's core selling point isn't just "speed." Diffusion language models work in a fundamentally different way from the autoregressive approach represented by the GPT series: autoregressive models generate token by token from left to right, while diffusion models produce text in parallel by progressively denoising noise. This paradigm has been proven in image generation, but transferring it to text generation has always been difficult when it comes to maintaining semantic coherence and long-range consistency. Inception has chosen to keep investing in this direction, and Mercury 3 is its latest-generation product.

The core mechanism of autoregressive models is token-by-token prediction: given the preceding text, the model outputs a probability distribution over the next token, samples from it, and feeds the result back as the next input, looping repeatedly. This makes the generation process inherently serial and constrained by memory bandwidth—a single forward pass produces only one token, and speed is hard to scale linearly. Diffusion models borrow the denoising idea from image generation: starting from random noise and iterating through multiple steps to jointly optimize the representation of the entire sequence, they can in theory process multiple positions in parallel within a single iteration. The differences between the two in parallelism, training objectives, and control granularity are the key background for understanding Mercury 3's speed claims.

Benchmark Comparison: Leading GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 Tests

According to a comparison table published by Inception itself, Mercury 3 leads GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 benchmarks. It should be noted that these results come from vendor self-testing; the choice of test sets, prompt design, and evaluation protocol are not detailed in the disclosed information, so they should be treated as reference data rather than independently verified conclusions.

In terms of positioning, both GPT-6 Luna and Claude Haiku 5.5 are lighter-weight models in their respective product lines that emphasize response speed. Mercury 3's choice of these as comparison targets suggests its intended use case is high-throughput, low-latency online inference, rather than complex reasoning tasks that chase the absolute ceiling. If the 1,068 tokens/second speed advantage holds under real workloads, it would mean significant cost and experience improvements for scenarios such as customer service, document processing, and code completion.

One detail worth noting: the theoretical advantage of diffusion models in parallel generation maps directly onto the inherent bottleneck of token-by-token decoding in autoregressive models. This is the underlying logic behind Inception's confidence in making speed its headline claim.

The difference between vendor self-testing and independent third-party evaluation deserves attention. Self-testing typically allows the vendor to choose prompt templates, the number of few-shot examples, decoding parameters (temperature, top-p), and scoring scripts—all of which can substantially affect final scores. Independent evaluations such as LMSYS Chatbot Arena, HELM, or the official leaderboards of academic benchmarks use unified protocols, blind testing, or human preference voting, making them more comparable. The industry convention is therefore: use vendor numbers to gauge magnitude and direction, but base procurement decisions on independent reproduction results.

Million-Token Context and the Compatibility Challenges of Diffusion Architectures

The 1 million token context is another metric worth unpacking. Long contexts place high demands on the memory footprint and computational complexity of attention mechanisms, and common industry approaches include sparse attention, chunked processing, and various approximation algorithms. For a diffusion LLM to keep the denoising process stable over long contexts is even harder—the propagation and convergence behavior of noise over long sequences differs from short-text generation.

Inception has not disclosed Mercury 3's actual effective recall performance at million-token contexts (for example, results on "needle in a haystack"-style tests), nor whether the speed figures were measured under full-context conditions. These two points directly determine whether the capability is engineering-ready or just a paper specification. For teams planning to use it for long-document analysis, we recommend running targeted validation before integration.

The "Needle in a Haystack" test is a common benchmark for evaluating the effective recall capability of long-context models: a specific piece of information (the "needle") is buried at an arbitrary position in a long text, and the model is then asked to accurately reproduce that information when answering a question. By testing repeatedly at different positions and text lengths, a heatmap of the model's recall accuracy can be drawn, revealing whether performance degrades in the middle of the sequence or under extremely long inputs. For models claiming million-scale contexts, this test is almost mandatory—there is often a clear gap between paper context length and real usable recall.

Practical Implications for Developers and Enterprise Model Selection

If Mercury 3's performance is confirmed by third-party evaluations, it could shift the selection logic in several areas.

Changes in inference cost structure. An order-of-magnitude speedup means a much larger volume of requests can be processed per unit of time, amortizing per-token costs on the same hardware. For teams deploying at scale, this is more attractive than a small lead in benchmark scores.

Reassessment of latency-sensitive applications. Real-time translation, voice assistants, interactive coding, and similar scenarios have long been constrained by generation speed. If the diffusion architecture can reliably output high-quality text, it would open up some previously uneconomical use cases.

Ecosystem compatibility remains a variable. New architectures typically face adaptation challenges in toolchains, fine-tuning methods, and deployment frameworks. Information is currently limited on whether Mercury 3 is compatible with mainstream inference frameworks and supports common fine-tuning paradigms, which will directly affect how quickly it can be adopted.

Where Caution Is Warranted

Leading on 8 of 12 benchmarks in vendor self-testing is still a long way from "comprehensively surpassing" competitors. What the four undisclosed losing metrics are and how large the gaps are also affect the judgment. The historical weaknesses of diffusion language models in factual accuracy and long-form logical consistency also need independent evaluation to confirm whether Mercury 3 has resolved them.

The speed figure of 1,068 tokens/second lacks a complete description of test conditions—batch size, hardware configuration, and output length all significantly affect results. Until third-party reproduction, a more prudent stance is to treat these numbers as directionally credible but unverified in magnitude.

For teams focused on inference efficiency, Mercury 3 is worth putting on the watch list; but for production-environment selection decisions, we recommend waiting for independent evaluations and hands-on integration testing before making a judgment.