Mercury 3: A Diffusion LLM That Scales 40x Without Slowing Down

Diffusion models have shone brightly in image generation, but in the language model arena they have long been seen as "non-mainstream." Mercury 3, newly released on Product Hunt, aims to change that. According to the official description, this is the most capable diffusion LLM to date, surpassing GPT-6 Luna and Claude Haiku 5.5 on multiple benchmarks while running at over 1000 tokens/s—4–8 times faster than its competitors.

A Differentiated Path for Diffusion LLMs

Nearly all mainstream large language models are based on the autoregressive architecture: generating token by token from left to right. Diffusion LLMs borrow from the image generation approach instead—iteratively denoising an entire block of noise until complete text "emerges." This paradigm naturally supports parallel decoding and, in theory, can substantially reduce inference latency. The release of Mercury 3 signals a key breakthrough in scaling diffusion architectures—the "40x scaling without slowing down" in the headline is a direct response to the historical difficulty of scaling up diffusion models.

The essential difference between autoregressive and parallel decoding

An autoregressive model can only determine one token per forward pass, because generating the nth token depends on the first n-1 tokens as conditional input. This sequential dependency makes inference impossible to parallelize along the time dimension, so latency grows linearly with output length. A diffusion LLM treats the entire token sequence as a whole: during training it learns the reverse process of gradually restoring original text from a noise-corrupted state, and during inference it starts from pure noise and runs a fixed number of denoising iterations to produce the complete output. Because each iteration operates on the entire sequence simultaneously, decoding is inherently parallel, and total latency depends on the number of denoising steps rather than output length—a fundamental difference from the token-by-token serialism of autoregressive models.

The accuracy–latency trade-off imposed by denoising steps

Inference quality in diffusion models is directly tied to the number of denoising steps: more steps usually yield better text quality but higher latency, while too few steps can produce semantic incoherence or repetition. Diffusion LLMs therefore have an adjustment knob that autoregressive models lack—developers can explicitly trade off quality against speed, a structural advantage in real-time scenarios requiring extremely low latency. But it also means that nominal throughput figures must be understood in the context of a specific step configuration.

The Trade-off Between Performance and Speed

Officially, Mercury 3 is claimed to surpass GPT-6 Luna and Claude Haiku 5.5 in capability while maintaining a 4–8x speed advantage. Note that all of these comparison figures come from the vendor's own statements and lack independent third-party reproduction. The Product Hunt page shows 94 votes and 2 comments, ranking 8th for the day—community interest is decent, but the comment sample is too small to form a consensus evaluation.

From an engineering perspective, a throughput of 1000+ tokens/s is highly attractive for real-time applications such as conversational agents, code completion, and batch document processing. If the figures hold, developers could support higher concurrency at the same hardware cost, or spend the saved compute on more complex inference pipelines.

The ambiguity of the tokens/s throughput metric

When measuring LLM inference speed, tokens/s has at least two common definitions: the per-token generation rate of a single request (which affects user-perceived time-to-first-token and output speed) and system-level aggregate throughput (which affects the concurrency supportable per unit of hardware cost). The two values can differ by an order of magnitude and depend heavily on batch size, sequence length, quantization precision, and hardware configuration. The parallel decoding of diffusion LLMs means they may maintain high utilization even at low batch sizes, unlike autoregressive models, which rely on large batches to amortize weight-reading costs. So before comparing claims like "4–8x faster," one must confirm that both sides were measured at the same batch size, output length, hardware, and precision—otherwise the numbers are not comparable.

The Team Behind It and Its Positioning

The maker is Volodymyr Kuleshov, and the product is categorized under three tags: API, Artificial Intelligence, and Developer Tools. This indicates that Mercury 3 is not an end-user chat application but an API-delivered model capability, aimed at development teams looking to replace their existing inference backend. For cost-sensitive or latency-sensitive scenarios, it is a noteworthy alternative.

A Rational View of the "Disruption" Narrative

For diffusion LLMs to truly challenge autoregressive models, they still need reproducible evaluation results on dimensions such as long-context consistency, instruction following, and tool calling. Mercury 3's 40x scaling and speed advantage are impressive engineering signals, but the benchmark-superiority claims still need community verification. For developers, the more pragmatic approach at this stage is to include it in small-scale comparative testing rather than directly replacing production pipelines.

(This article is based on public information from Product Hunt; performance data is subject to official release.)

Why long-context consistency is difficult for diffusion LLMs

An autoregressive model can attend to all previously generated content when generating each token, and long-range dependencies are maintained position by position through the attention mechanism under a causal mask. A diffusion LLM, by contrast, rewrites the entire sequence simultaneously across multiple denoising rounds; in early iterations, consistency constraints between distant tokens are weak, making it prone to referent drift, factual contradictions, or structural breakdown—especially in long outputs of thousands of tokens or more. Instruction following and tool calling likewise depend on strict sequence-structure constraints (such as JSON syntax and function signatures), and the globally parallel updates of diffusion decoding do not inherently guarantee that these constraints are satisfied position by position, usually requiring additional constrained decoding or post-processing correction. This is exactly why diffusion LLMs cannot be evaluated on short question-answering benchmarks alone.

One-line Summary

Mercury 3 challenges the autoregressive mainstream with parallel denoising, claiming superiority in both speed and capability—pending verification.

Paragraph Summary

Mercury 3 is a diffusion LLM released on Product Hunt. Officially described as the most capable diffusion language model to date, it is claimed to surpass GPT-6 Luna and Claude Haiku 5.5 on multiple benchmarks, reaching over 1000 tokens/s—4–8 times faster than competitors. Unlike token-by-token autoregressive models, diffusion LLMs iteratively denoise an entire block of noise, naturally supporting parallel decoding and, in theory, significantly reducing inference latency; the "40x scaling without slowing down" in the headline points to a key breakthrough in scaling diffusion architectures. The product was released by Volodymyr Kuleshov as an API for developers, suited to latency- and cost-sensitive scenarios. However, all performance comparisons come from the vendor's own statements and lack third-party reproduction, and with only 94 votes and 2 comments on Product Hunt, community consensus has yet to form. The article recommends including it in small-scale comparative testing rather than directly replacing production pipelines.