Mercury 3 Hits 1,068 Tokens Per Second as Diffusion Language Models Begin to Scale

Inception Labs released Mercury 3 on October 19, 2026, a production diffusion language model generating text at 1,068 tokens per second with a one-million-token context window. Company-reported benchmarks put it ahead of GPT-6 Luna and Claude Haiku 5.5 in eight of twelve tests, but none of those results have been independently reproduced.

Buried in the launch materials is a notable detail: Inception used Mercury 3 agent swarms during the model's own training. CEO Stefano Ermon told The Deep View newsletter that this recursive process added approximately 16 percent more throughput on top of gains from manual optimization — the model being trained contributed to training itself.

Diffusion Generation Removes the Autoregressive Bottleneck

Standard language models generate tokens one at a time: each position waits for the previous one to complete. Diffusion language models invert that process. Mercury 3 begins with a fully masked sequence — every token hidden — and iteratively unmasks positions across multiple refinement passes using masked discrete diffusion, resolving many tokens in parallel per step. Because each step covers the whole sequence simultaneously, throughput is not bounded by the serial token-prediction bottleneck that limits autoregressive models.

The masking approach differs from image diffusion. Language tokens are categorical, not continuous, so Mercury uses discrete token masking (similar to BERT's masked language modeling) rather than Gaussian noise. Bidirectional attention allows every position to attend to full context from both directions — unlike the causal, forward-only attention of GPT-style models, which can only look back.

Inception's founders come directly from diffusion model research. CEO Stefano Ermon is a Stanford professor whose work on score-based generative models underlies Stable Diffusion and related image systems; applying that lineage to language is the company's founding purpose.

The Benchmark Table Shows a More Mixed Picture

Mercury 3 wins eight of twelve tests in Inception's benchmark comparison, all run at medium inference settings. The largest gap is instruction following: IFBench gives Mercury 3 a score of 76.84, roughly 15 points above GPT-6 Luna and Claude Haiku 5.5, which score approximately 61.8 and 60.3 respectively. On SWE-Bench Atlas, which tests autonomous agentic coding with real tool use, Mercury 3 reaches 51.08 percent against 33.33 percent for Claude Haiku 5.5 and 32.26 percent for GPT-6 Luna.

The four tests Mercury 3 loses are significant. Claude Haiku 5.5 wins DeepSWE (longer, more complex coding tasks) and long-context retrieval. GPT-6 Luna wins two tests from Artificial Analysis. Individual scores for the six remaining Mercury 3 wins are not publicly detailed. Teams building code-heavy agent pipelines or long-document retrieval applications will find Haiku 5.5 performing better on exactly those tasks.

All numbers are company-reported. Inception ran its own benchmarks, competitor throughput figures are Inception-cited rather than drawn from OpenAI or Anthropic documentation, and no external reproduction exists.

Scaling Evidence and a Self-Improving Training Loop

The most architecturally significant claim at launch is not the throughput number: it is the scaling finding. Mercury 3 carries more than 40 times the parameters of Mercury 1 while generating at the same 1,068 tokens per second. Inception presents this as evidence that diffusion models obey the same scaling laws as autoregressive transformers — larger models improve without proportional latency cost. Google's DiffusionGemma, released in June 2026, independently validates the direction: the company reported up to 4× throughput versus autoregressive Gemma 4, while noting a slight quality tradeoff.

Mercury 2.5, the previous model, ran at 1,107 tokens per second with a 260,000-token context window. Mercury 3 is marginally slower but accepts four times longer contexts. The 1 million-token window enables workloads — full codebases, long document collections — that previously required retrieval augmentation.

Ermon's agent swarm training claim adds another dimension. An earlier version of Mercury 3 was deployed as a swarm to improve the model's own training pipeline, contributing the roughly 16 percent throughput gain. Most commercial AI labs describe model improvements in terms of architecture and data decisions; recursive agent-based self-optimization is not standard practice in publicly announced development.

Where Latency Compounds and What Partners Report

The strongest use case for Mercury 3 is AI agent pipelines where model latency accumulates across many sequential calls. Each step in an autonomous workflow waits for a model response; 130 tokens per second versus 1,068 tokens per second creates a material latency difference across five or ten consecutive calls.

Octen AI, which builds search for smart glasses, told Inception that language generation was the pipeline bottleneck: the search retrieval itself took about 80 milliseconds, but total response time extended to five seconds. With Mercury 3, the company reports a median response of 0.9 seconds and a full-process time of 2.2 seconds. Education company General Learning reports delivering student hints and corrections in approximately one second. TinyFish AI says Mercury 3 solved 25 of 50 real website navigation tasks, outperforming GPT-6.1-Sol — no score is given for the OpenAI model. All partner results are company-sourced.

Pricing, Availability, and What Comes Next

Mercury 3 launched at $0.10 input and $0.35 output per million tokens — roughly half the standard rate. Inception has not disclosed when that discount ends. At regular rates of $0.20 input and $0.70 output, Mercury 3 charges 40 percent more per output token than GPT-6 Luna and Claude Haiku 5.5, both priced at $0.50 per million output tokens for shorter contexts. For very long inputs, Mercury 3's flat per-token rate may be cheaper than rivals' tiered pricing, but cost modeling for typical output-heavy agent workloads will need to account for that post-discount output premium.

The model is available through Inception's API, OpenRouter, Baseten, and Vercel with an OpenAI-compatible interface, reducing integration friction for teams switching from GPT-based pipelines.

The development that would most change Mercury 3's market position is also the one Inception cannot produce: third-party benchmark reproduction. The 15-point IFBench lead and the instruction-following and agentic-coding advantages all derive from Inception's own evaluation. Independent results will determine whether those advantages replicate under equivalent conditions — and whether the model trained partly by its own agent swarms performs as the benchmark table suggests on tasks it has never seen from outside the company.