Mercury 3: Inception's Diffusion LLM Runs at 1,068 Tokens/Second With a 1M-Token Context
Inception Labs has released Mercury 3, which it calls the most capable diffusion-based LLM released to date. The company claims Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 on benchmarks while running 4–8x faster at 1,000+ tokens per second, and it presents the model as proof that text diffusion scales the same way autoregressive language models do.
Mercury 3 at a glance
Mercury 3 is a diffusion LLM served at 1,068 tokens per second on NVIDIA GPUs, with a 1M-token context window and launch pricing below the speed-optimized frontier models it compares itself against.
| Spec | Mercury 3 |
|---|---|
| Model type | Diffusion-based LLM |
| Speed | 1,068 tokens per second on NVIDIA GPUs |
| Context window | 1M tokens (up from 260K for Mercury 2.5) |
| Standard price | $0.20 per 1M input tokens, $0.70 per 1M output tokens |
| Launch price | $0.10 per 1M input tokens, $0.35 per 1M output tokens (50% off) |
| Cached tokens | 90% off the input price |
| Availability | Inception API Platform, OpenRouter, Baseten, Vercel |
| API compatibility | OpenAI API compatible |
| Free tier | 100 million free tokens on the Inception platform |
Inception states that Mercury 3 is "on par or better than frontier speed-optimized models like GPT-6 Luna and Claude Haiku 5.5" on quality benchmarks. These are the company's own benchmark results; the announcement does not publish the underlying evaluation methodology in the scraped text.
The scaling claim: diffusion scales like autoregressive models
The central argument of the announcement is that diffusion LLMs follow scaling laws, not just autoregressive ones. Inception says Mercury 3 is more than 40x larger in total parameters than Mercury 1, and that intelligence across a range of standard benchmarks rose 3x from Mercury 1 (mini) to Mercury 3.
When Mercury 1 launched, Inception framed it as "the first proof point that diffusion LLMs could be built for commercial applications." The open question at the time, according to the company, was whether diffusion could scale to compete with autoregressive models on complex use cases. Inception's answer in the Mercury 3 announcement is direct: "they can."
The company's stated conclusion is that "the scaling laws that drove progress in autoregressive models hold for diffusion too."
Production deployments and partner results
Inception says early-access partners are already running Mercury 3 in latency-sensitive applications including browser use agents, interactive education, real-time search agents, and customer support. The announcement includes named partner testimonials with specific performance numbers.
Browser use agents: TinyFish AI
Inception argues that agents wait on the model at every step, so small delays compound. It reports running Mercury 3 inside its own production web-agent harness against frontier models on 50 live Mind2Web navigation tasks. Mercury 3 completed 50% of them, which Inception says is significantly more than GPT-6.1-Sol, using the same token volume and the same number of agent steps as a frontier model, at a fraction of the per-token cost.
Shuhao Zhang, cofounder of TinyFish AI, is quoted saying agents "finish sooner and have time to try more approaches before answering."
Interactive education: RevisionDojo
For tutoring and grading workloads, Inception claims Mercury 3 responds in near real time. Michael Tong, co-founder and CTO of General Learning (aka RevisionDojo), said the model "allowed us to give every student instant feedback without giving up reasoning," with hints, marked answers, and definitions returning "in about a second," at quality "as good as or better than the models we ran before." He described it as "a drop-in swap for our existing stack."
Real-time search agents: Octen AI
Inception states that search applications feel unresponsive after just a few seconds of delay. In Octen AI's Sense API tests, Octen search took about 80ms and model generation was the bottleneck. Kuan Zou, founder and CEO of Octen AI, said Mercury 3 "delivered the best balance of speed and quality among the fast models we tested," with a median generation time of 0.9 seconds and a median end-to-end response time of 2.2 seconds.
Customer support: Lorikeet
Lorikeet, which builds concierge-style support across voice, chat, and email for regulated businesses, reported that in early testing Mercury 3 "beat other models on latency while also performing higher on evals and in production," according to Michael Gribben, a member of technical staff.
Mercury Voice and Mercury Decide
Mercury 3 is the flagship of a wider Inception product portfolio aimed at real-time AI. Two other products are named in the announcement:
- Mercury Voice is customized for extreme low-latency voice applications and for minimizing time to first answer token (TTFAT), not just end-to-end latency. Inception says it is already powering agents at Audivi AI, Altur, and OpenCall.
- Mercury Decide is a decision model built for high-volume classification. Inception claims it outperforms Jev on benchmarks including JevBench and LegalBench.
Availability and pricing
Mercury 3 is available today on the Inception API Platform and through OpenRouter, Baseten, and Vercel. It is OpenAI API compatible, so Inception says it drops into existing stacks with no rewrites. Enterprise deployments support dedicated capacity, autoscaling, compliance controls, and configurable data retention.
Inception is offering 100 million free tokens to try the API, plus a chat interface at chat.inceptionlabs.ai.
What Hacker News commenters said
Discussion on Hacker News (176 points, 71 comments) mixed hands-on reports with skepticism about benchmark comparisons and questions about open weights.
Hands-on impressions were largely positive
One commenter, Escapado, reported building a macOS Swift UI app for OCR and translation with Mercury 3 in about 45 minutes of prompting, using 26 million tokens (1.3M uncached) at around 160 tokens per second:
"I made it come up with a plan for this tool and had it implement the thing end to end in about 8 or 9 steps. It originally messed up the coordinate calculation but quickly corrected that when I pointed it out and other than that it one-shotted all the other steps."
GodelNumbering said the model "looks noticeably more capable than previous gen" after asking it to build an HTML renderer for UML diagrams, and noted it picked "different and less painful-to-eyes color schemes for each theme." UncleOxidant, after trying it on a repo, wrote: "Pleasantly surprised." moonu called it "a definite improvement over 2.5" but wished the models were multimodal.
Skepticism about benchmark comparisons
Not every hands-on report matched the marketing. snake_doc said that on OpenRouter, "deepseek v4.1 flash via together/baseten/fireworks still beat mercury 3 on latency and intelligence," adding that "Ttft is quite slow with mercury." vatsachak reported it was "not as fast as GPT 6.1-Sol on my favorite toy math puzzle," speculating that "intelligence consumes less tokens."
jjaskulowski raised a methodological objection to the head-to-head framing:
"Any comparision would make sense only if you made them - Mercury and Luna - the same training data and then the same post process, alignment. It's nice it's fast. I support it. Just saying I don't know how to judge that."
The Google Gemini Diffusion question
Simon Willison noted that Inception is not the only lab that has shown diffusion text generation:
"I still don't understand why Google did effectively nothing with their Gemini Diffusion demo from last year. It was really cool, and really fast, and just went quiet. They did release a Diffusion Gemma a few months ago but I've not seen much of a splash from that one, either."
dumberquestions pushed back on the idea that Inception has a durable lead: "I'd be surprised if major labs aren't doing any diffusion experiments, and if they are, it would also be surprising if another lab is ahead of them in diffusion given just how much more compute they have at their disposal."
Cost, open weights, and unanswered questions
2001zhaozhao asked how Inception achieved the pricing, calling Mercury 3 "a no-brainer against Haiku 5.5 and GPT-6 Luna" if the benchmarks hold up, and noting that "Haiku 5.5 becomes 5x more expensive past 100k context, and this model goes up to 1M for the same price." The same commenter added that the price increases 2x after a period post launch but remains far below Haiku's pricing above 100k requests. moecables echoed the concern: "I'm wondering if it'll still be useful once the discount is lifted?"
SillyUsername was disappointed on the open-weights front: "Damnit, I've been waiting on an open weights model like this one, still no dice." UncleOxidant similarly hoped "we see more open weight diffusion LLMs (currently there's diffusion gemma)."
Other commenters raised questions the announcement does not answer. loufe asked what the actual model size is and how training differed from a traditional LLM. GodelNumbering asked how much code a diffusion model needs to read for incremental changes in a codebase and how well it handles tool calls. fraywing asked whether these models are "harder to reason about given they produce tokens non-linearly." davedx wondered how competitive Mercury 3 would be against decision models like Jev, given that decision workloads want maximum throughput. nkoren noted that a comment from Inception co-founder volodia appeared flagged or dead on the thread.
Caveats
The performance, quality, and cost figures in this article come from Inception Labs' own announcement and from named partner testimonials. The announcement does not disclose model parameter counts, training data, or full benchmark methodology. Independent commenter reports on Hacker News were mixed, with some users reporting strong results and others reporting that competing models such as DeepSeek v4.1 Flash matched or beat Mercury 3 on latency in third-party serving environments.