Inception Mercury 3 Review: The First Diffusion LLM to Go Big — What 1,068 Tokens/Second Actually Means for You
On October 19, 2026, Inception — a Redwood City startup backed by Menlo Ventures, Mayfield, Microsoft's M12, and individual investors including Andrew Ng and Andrej Karpathy — released Mercury 3, the newest version of the only commercial "diffusion" large language model line. Inception says Mercury 3 generates text at 1,068 tokens per second on NVIDIA GPUs, outperforms OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 on a range of benchmarks while running four to eight times faster, and carries a context window of one million tokens. Launch pricing is aggressive: $0.10 per million input tokens and $0.35 per million output tokens.
Those are striking claims. They are also, for now, mostly Inception's claims. In this review we separate what is verified from what is vendor-reported, explain what a diffusion LLM actually is (in plain language), look at what early testers and developers are really saying, and cover the questions the press release doesn't answer — including the deadline quietly attached to this launch for existing Mercury users.
What Is a Diffusion LLM? (The Simple Explanation)
Almost every AI model you know — ChatGPT, Claude, Gemini — is autoregressive. It writes text the way a person types: one token (roughly, one word-piece) at a time, left to right, where each new token depends on everything written before it. It is accurate, but it is inherently sequential: token number 500 cannot be written until tokens 1 through 499 exist.
A diffusion LLM (dLLM) works more like an image generator such as Stable Diffusion. Instead of typing one word at a time, the model starts with a rough, noisy draft of the whole answer and refines it in parallel passes — fixing many tokens at once, the way a photo sharpens from blur to clarity. Because large chunks of the answer are produced simultaneously, the output arrives much faster.
The trade-off the industry has always worried about: parallel refinement has historically meant weaker reasoning and sloppier long answers than step-by-step generation. Inception's entire pitch for Mercury 3 is that this trade-off is now closed — that diffusion can scale in intelligence the same way autoregressive models did. Co-founder Stefano Ermon put it this way at launch: "Since we launched Mercury 1, the question we've heard most is whether diffusion can scale. Mercury 3 is over 40 times larger, significantly more capable, and still runs at over 1,000 tokens per second."
Mercury 3 is Inception's fourth release in sixteen months: Mercury 1 (June 2025) proved the concept, Mercury 2 added reasoning, Mercury 2.5 added tunable reasoning and a 260,000-token context window, and Mercury 3 jumps to one million tokens. Mercury 2 was also named to TIME's Best Inventions of 2026 list, announced October 15.
Mercury 3: The Verified Specs
| Specification | Detail | Source |
|---|---|---|
| Release date | October 19, 2026 | Inception press release |
| Model type | Diffusion large language model (dLLM) | Inception |
| Generation speed | 1,068 tokens/second on NVIDIA GPUs (vendor-reported) | Inception |
| Context window | 1,000,000 tokens (up from 260,000 in Mercury 2.5) | Inception |
| Model size | "More than 40× the total parameters of Mercury 1" — absolute parameter count not disclosed | Inception; noted by D-CSIL briefing |
| Input price (list) | $0.20 per million tokens | Inception models page |
| Output price (list) | $0.70 per million tokens | Inception models page |
| Launch price (50% off) | $0.10 input / $0.35 output per million tokens; cached input $0.01 per million | Inception models page |
| API model ID | mercury-3, OpenAI-API compatible | Inception docs |
| Access | Inception API Platform, OpenRouter, Baseten, Vercel | Inception |
| Free tier | 100 million free tokens for new platform accounts, no payment details required | Inception docs |
| Weights | Closed — API access only, no open-weight release | Observed; no official model repo exists |
What Early Testers Found
The most useful early signal comes from The Neuron's hands-on review, published within hours of launch. Their spot checks were genuinely encouraging: asked to build a complete single-file HTML dungeon crawler game, Mercury 3 produced working JavaScript that passed a syntax check and survived 2,000 randomized map-generation tests with zero failures. On a classic logic puzzle, it answered correctly and completely for a total cost of $0.00014. On hallucination resistance, it correctly refused to invent details about a fictitious NASA mission — although it then introduced an unrelated factual error in its supporting details, a reminder that rejecting a false premise is not the same as being right about everything else.
The Neuron was equally clear about the limits of its own testing: these were editorial spot checks, not a controlled benchmark study, and the reviewer did not independently reproduce the 1,068 tokens/second figure, noting that "company-reported peak throughput and what any given user sees through a third-party provider can differ considerably." Other early coverage has been more mixed on coding, with at least one early take reporting Mercury 3 trailing Claude Haiku 5.5 on coding tasks. The honest summary on day one: promising, fast, cheap — and not yet independently benchmarked.
Inception's own production evidence is worth knowing, with the caveat that Inception ran it. The company says it tested Mercury 3 inside its own web-agent harness on 50 live Mind2Web navigation tasks and completed 50% of them — a completion count it describes as significantly higher than GPT-6.1-Sol's, at the same token volume and step count. Launch testimonials from early customers are consistent on one theme — latency people can feel: tutoring company RevisionDojo reports student feedback returning in about a second; smart-glasses maker Octen AI measured a median generation time of 0.9 seconds in its Sense API tests; customer-service builder Lorikeet reports Mercury 3 beating other models on latency in early testing while scoring higher on its evaluations.
What Developers Are Actually Asking
Digging through Hacker News threads and GitHub activity around the Mercury line, the developer conversation is more cautious than the launch headlines — and more useful. The recurring themes:
- Tool calling and structured output. The single most repeated question from developers who evaluated Mercury 2 is whether tool calling and structured output are "truly consistently possible" at the level of Anthropic's Haiku — because that, more than raw speed, decides whether the model can run real agents.
- Agentic coding in custom harnesses. One team that tested Mercury 2.5 Preview reported it performed worse on their problems when plugged into a custom coding harness, concluding Inception "had not figured out general purpose tool use and agentic coding" — while adding that if they do, "the speed and cost will enable" a lot of real-time applications.
- Reliability basics. A developer testing via the chat site hit errors and discovered there was no public status page to check — a small thing that matters a lot in production.
- Will real-world speed match the headline? Independent briefings (D-CSIL) stress that the benchmark and speed numbers are Inception's own, and that builders need to evaluate the model on their own tasks and serving conditions.
On GitHub, the ecosystem is community-built rather than official: there is a third-party Python SDK for the Mercury API, a plan/execute coding harness with a benchmark written specifically for Mercury 2.5, several showcase apps, and an independent API profile of Inception that was updated on launch day. Hundreds of code references to Mercury 3 appeared within the first day — developers are clearly trying it. But there is no official Inception model repository, because there are no weights to publish.
What the Press Release Doesn't Say
- The exact model size. "40× larger than Mercury 1" is a ratio, not a number. Inception has not disclosed Mercury 3's parameter count.
- Independent verification. Every benchmark win and the 1,068 tok/s speed figure come from Inception. No third party had reproduced them at publication time.
- When the 50% launch discount ends. The pricing pages state no end date. Co-founder Stefano Ermon has said he hopes the lower price could become permanent — a hope, not a commitment. Budget at the list price ($0.20/$0.70) to be safe.
- The migration deadline. Mercury 2 and Mercury 2.5 are closed to new users as of this launch. Existing customers keep access only until November 15, 2026 (enterprise agreements excepted). If you built on Mercury 2.x, your migration clock is already running.
- Open weights. Mercury 3 is API-only. If your use case requires self-hosting, this model is not an option at any price.
- The tool-use question. The launch materials lean on testimonials and Inception's own agent harness. The community's open question from the Mercury 2 era — dependable tool calling and task adherence versus Haiku-class models — is not settled by anything published so far.
Who Should Try Mercury 3 (and Who Should Wait)
Try it now if: your application is latency-sensitive — voice and chat agents, live tutoring or feedback tools, real-time search agents, customer support, browser-use agents. This is where partners are already running it in production, and where a 4–8× speed advantage (if it holds on your workload) changes what is possible. The 100 million free tokens, with no payment details required, make a serious evaluation genuinely free.
Test carefully first if: you depend on tool calling, structured output, or multi-step agentic coding. Run Mercury 3 against your current Haiku-class setup on your own tasks before switching anything that matters — that is exactly what independent reviewers recommend.
Wait if: you need open weights, self-hosting, or independently verified benchmarks before adopting. None of those exist yet.
How to access it: sign up on the Inception API Platform (model ID mercury-3, OpenAI-compatible endpoint, with official Python and TypeScript SDKs and support for LiteLLM, LangChain, and the Vercel AI SDK), or call it through OpenRouter, Baseten, or Vercel if you already work there. Because the API mirrors OpenAI's format, most teams can point existing code at it with minimal changes — RevisionDojo described the switch as a "drop-in swap."
Related reading on CloudEmythos:
FAQ
What is a diffusion LLM, in simple terms?
A normal AI model writes one token at a time, left to right. A diffusion LLM drafts the whole answer at once and refines it in parallel passes — like a photo coming into focus. That parallelism is why Mercury 3 can generate text so much faster than conventional models of similar quality.
Is Mercury 3 really 1,068 tokens per second? Will I see that speed?
1,068 tokens/second is Inception's reported figure on NVIDIA GPUs. Early hands-on reviewers confirm the API feels very fast but have not independently reproduced the number, and your actual throughput will vary by provider (Inception direct vs. OpenRouter vs. Baseten), prompt length, and load. Treat it as a vendor-reported peak, and measure it on your own workload.
How much does Mercury 3 cost — and is the 50% discount permanent?
List pricing is $0.20 per million input tokens and $0.70 per million output tokens. Launch pricing is half that — $0.10 / $0.35 — with cached input at $0.01 per million. Inception has published no end date for the discount, and its co-founder has said only that he hopes the lower price could become permanent. New accounts also get 100 million free tokens with no payment details required.
Is Mercury 3 good for coding and AI agents?
Early evidence is mixed, and this is the model's biggest open question. One hands-on review found its one-shot code generation impressive and reliable across thousands of automated checks, while other early takes report it trailing Claude Haiku 5.5 on coding. Developers evaluating earlier Mercury versions flagged tool calling, structured output, and custom-harness agent work as weak points. Inception counters with its own web-agent results (50% completion on 50 live Mind2Web tasks). Verdict: benchmark it on your tasks before migrating.
Can I download or self-host Mercury 3?
No. Mercury 3 is a closed, API-only model — there are no open weights and no official model repository. You can access it through Inception's platform or through OpenRouter, Baseten, and Vercel.
Verdict
Mercury 3 is the most credible challenge yet to the assumption that AI models must choose between fast and smart. The architecture is genuinely different, the pricing is genuinely aggressive, the free tier is genuinely generous, and the launch-day evidence — partner production use, one strong independent spot check, fast community uptake on GitHub — is better than a bare press release.
But the discipline we apply to every launch applies here too: the headline numbers are the vendor's, the parameter count is undisclosed, the discount has no end date or guarantee, and the developer community's core question — can it call tools and stick to tasks like a Haiku-class model? — remains open until independent testing settles it. If speed is your bottleneck, Mercury 3 deserves a serious, free evaluation this week. Just run that evaluation on your own tasks, and if you're still on Mercury 2 or 2.5, put November 15, 2026 on your calendar.
Sources: Inception launch press release (Business Wire, October 19, 2026); Inception models page and API documentation via Unite.AI and D-CSIL briefings; The Neuron hands-on review; The Deep View; Hacker News developer discussions of Mercury 2/2.5; GitHub community projects. All benchmark and speed figures are Inception-reported unless noted as independently observed.
No comments:
Post a Comment