Inception's Mercury 3 claims diffusion LLMs scale, with 1,000+ tokens per second

Inception releases Mercury 3 with a speed-first pitch

Inception has launched Mercury 3, which it describes as the most capable diffusion-based language model released to date. According to the company, the model matches or beats speed-optimized frontier models GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running four to eight times faster than those rivals, at more than 1,000 tokens per second. Inception measured 1,068 tokens per second on NVIDIA GPUs.

The model also ships with a one-million-token context window, up from 260,000 in Mercury 2.5. List pricing is $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount bringing that to $0.10 and $0.35; cached tokens are billed at 90% off the input price. Inception is offering 100 million free tokens for developers to try the API.

Inception's core argument: diffusion scales

Diffusion language models work differently from the autoregressive models that dominate the market. Rather than emitting one token after another, they generate text by refining it in parallel passes, an approach Vercel's changelog credits for the throughput figures.

The deeper claim in the announcement concerns scale. When Inception launched Mercury 1, the open question in the field was whether diffusion models could grow to compete with autoregressive systems on complex tasks. Inception says Mercury 3 settles it: the model is more than 40 times larger than Mercury 1 in total parameters, benchmark performance has tripled from Mercury 1 mini to Mercury 3, and generation speed has held steady. The company's conclusion is that the scaling dynamics that drove progress in autoregressive models apply to diffusion as well.

What early partners report

Inception says early-access partners are already running Mercury 3 in latency-sensitive products: browser automation agents, interactive education, real-time search and customer support.

The company published one concrete result: inside its production web-agent harness, on 50 live Mind2Web navigation tasks, Mercury 3 completed 50% of them, which Inception describes as significantly more than GPT-6.1-Sol, while using the same token volume and number of agent steps as a frontier model at a fraction of the per-token cost.

Customer statements quoted in the announcement support the latency story. Michael Tong, co-founder and CTO of General Learning, the company behind RevisionDojo, says hints, marked answers and definitions now come back in about a second. Kuan Zou, founder and CEO of Octen AI, reports a median generation time of 0.9 seconds and 2.2 seconds end to end in his company's smart-glasses search tests. Lorikeet's Michael Gribben says Mercury 3 beat other models on latency while also scoring higher on evaluations and in production.

Portfolio and availability

Mercury 3 is the flagship in a broader lineup. Mercury Voice is tuned for voice applications and minimizes time to first answer token, and Inception says it already powers agents at Audivi AI, Altur and OpenCall. Mercury Decide, a decision model aimed at high-volume classification, is claimed to outperform Jev on benchmarks including JevBench and LegalBench.

The main model is available on the Inception API Platform as well as OpenRouter, Baseten and Vercel, and it is OpenAI API compatible, so existing stacks need no rewrites. Enterprise deployments add dedicated capacity, autoscaling, compliance controls and configurable data retention.

Vercel's changelog confirms availability through its AI Gateway at the discounted launch price, using the model name inception/mercury-3, reachable from tools like Claude Code, Codex and Cursor after running the gateway setup in the Vercel CLI. Vercel passes provider pricing through with no markup and no platform fee on inference, including bring-your-own-key requests.

Why it matters

Nearly every widely deployed LLM is autoregressive, and diffusion-based text generation has mostly lived in research. If Inception's numbers hold up outside its own benchmarks, Mercury 3 is the first commercial-scale evidence that diffusion models follow the same scaling curves while removing the token-by-token latency bottleneck. That combination targets a real gap: agents, voice interfaces and tutoring all compound small delays at every step, and a model that reasons at frontier-adjacent quality but answers in under a second changes what those products can feel like. The caveats are the usual ones for a launch announcement — the comparisons come from the vendor, and independent evaluations have not yet weighed in — but the free-token trial and multi-platform availability make it cheap for developers to test the claims directly.