Inception Claims 1,068 Tokens Per Second With Mercury 3 Diffusion LLM

Inception says Mercury 3, its new diffusion language model, generates text at 1,068 tokens per second on NVIDIA GPUs — roughly eight times faster than GPT-6 Luna and four times faster than Claude Haiku 5.5 by the company's own measurement. Inception co-founder Stefano Ermon announced the release on October 19, 2026, writing: "Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec."

The speed claim is Inception's own, measured on its own served infrastructure. The benchmark wins come from Inception's own table, run at medium settings. Every headline number in this launch, in other words, comes from the vendor selling the model.

What is Mercury 3?

Mercury 3 is Inception's fourth Mercury release, following Mercury 1 (launched in 2025), Mercury 2 and Mercury 2.5, which added tunable reasoning and a 260K context window. Inception says the new model is more than 40 times larger than Mercury 1 (mini), though it discloses no absolute parameter count. Context window is 1M tokens, matching Haiku 5.5 and sitting just under Luna's 1.05M.

The model is API-only. There are no open weights and no quantized builds, and Inception does not disclose which NVIDIA GPUs serve it. Maximum output length and input modalities are also undisclosed — both rivals publish a 128K output cap and accept text and image.

How does a diffusion LLM work, and does it scale?

Conventional autoregressive models write one token at a time; each token waits for the previous one. A diffusion LLM (a model that generates by refining a rough draft of the entire answer across a few parallel passes) removes much of that sequential bottleneck. Fewer sequential steps mean faster output on identical hardware.

The open research question was whether that approach survives scale. Inception claims it does: moving from Mercury 1 to more than 40x the parameters, average benchmark quality rose roughly 3x while speed stayed above 1,000 tokens per second. The company plots this as a quality-versus-size curve across five Mercury generations. Those curves are drawn from the company's own runs.

Takeaway: the vendor claims diffusion holds its speed advantage as models grow, but no independent lab has yet checked the plot.

How does Mercury 3 perform on benchmarks?

All scores come from Inception's table, with every model run at medium settings. Mercury 3 leads GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests.

The largest gains land on instruction following and agentic coding:

• IFBench: 76.84 — 15 points above GPT-6 Luna (61.82); Haiku 5.5 scores 54.65.
• SWE-Bench Atlas: 51.08 — against 33.33 for Luna and 32.26 for Haiku 5.5, a gap of about 18 points.
• GPQA Diamond: 89.23 — versus 86.36 for Luna and 84.51 for Haiku 5.5.

The losses matter as much. Haiku 5.5 wins DeepSWE v1 (51.62 vs 46.31) and MRCR v2 (78.40 vs 75.50). GPT-6 Luna edges ahead on AA-Omniscience-Public (46.94 vs 46.44) and AA-LCR v1.1 (78.30 vs 77.67). Long-context retrieval and longer coding tasks remain the model's weak spots — a meaningful caveat for a model whose headline spec is a 1M-token context window.

What do early users report?

Inception quotes four early-access partners. These are customer statements, not independent evaluations.

• TinyFish ran 50 live Mind2Web web-navigation tasks; Mercury 3 completed 50% of them. Inception says that beat GPT-6.1-Sol but provides no Sol number.
• Octen AI reports 0.9 seconds median generation and 2.2 seconds end to end for smart-glasses search.
• RevisionDojo says hints and marked answers return in about a second.
• Lorikeet says Mercury 3 beat other models on latency and evals for support agents.

The TinyFish comparison is the only one with a task count attached, and its baseline is missing. The rest are latency impressions without published methodology.

How can developers use it, and what does it cost?

Mercury 3 is live on the Inception API Platform, OpenRouter, Baseten and Vercel. The API is OpenAI-compatible; the docs use the model string mercury-3 and a reasoning_effort parameter.

Pricing at list is $0.20 per 1M input tokens and $0.70 per 1M output tokens. The launch discount halves that to $0.10 and $0.35 — at which point it matches GPT-6 Luna and Haiku 5.5 on input while undercutting them on output ($0.50 for both rivals). Cached input is 90% off, and new accounts receive 100M free tokens. Enterprise plans add dedicated capacity, autoscaling and configurable data retention.

Inception also ships two siblings: Mercury Voice, targeting time to first answer token in voice agents, and Mercury Decide, a classification model.

Why it matters

Diffusion language models have been a research curiosity since Inception shipped Mercury 1 in 2025; a 40x scale-up with intact throughput is the strongest evidence yet that the architecture can compete at frontier quality. But the entire case rests on vendor-run benchmarks, undisclosed parameters and undisclosed GPUs. Independent verification — or a refusal to publish one — will determine whether 1,068 tokens per second is an engineering milestone or a marketing table.