--- title: Mercury 3: Diffusion at scale | daily.dev description: Inception Labs launched Mercury 3, a diffusion-based LLM it claims is the most capable released to date, outperforming GPT-6 Luna and Claude Haiku 5.5 while... canonical: https://daily.dev/posts/mercury-3-diffusion-at-scale-lnpntxxhe twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Mercury 3: Diffusion at scale | daily.dev og:description: Inception Labs launched Mercury 3, a diffusion-based LLM it claims is the most capable released to date, outperforming GPT-6 Luna and Claude Haiku 5.5 while... og:url: https://daily.dev/posts/mercury-3-diffusion-at-scale-lnpntxxhe og:image: https://api.daily.dev/og/posts/lnPNTXXhe.png og:image:alt: Mercury 3: Diffusion at scale og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Mercury 3: Diffusion at scale **[Hacker News](https://daily.dev/sources/hn)** · 5 min read · 3 upvotes · 0 comments ## Summary Inception Labs launched Mercury 3, a diffusion-based LLM it claims is the most capable released to date, outperforming GPT-6 Luna and Claude Haiku 5.5 while generating over 1,000 tokens per second (4-8x faster than comparable models). It's over 40x larger than Mercury 1, supports a 1M token context window (up from 260K), and costs $0.20/$0.70 per million input/output tokens (50% off at launch). Early-access partners including TinyFish AI, RevisionDojo, Octen AI, and Lorikeet report using it for browser agents, tutoring, search, and customer support due to its low latency. The company also introduced Mercury Voice and Mercury Decide as part of a broader real-time AI product portfolio. The model is available via the Inception API, OpenRouter, Baseten, and Vercel, and is OpenAI API compatible. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How fast is Inception's Mercury 3 diffusion model compared to GPT-6 Luna and Claude Haiku 5.5? Mercury 3 runs 4-8x faster than GPT-6 Luna and Claude Haiku 5.5, generating over 1,000 tokens per second (1,068 tokens per second on NVIDIA GPUs) while matching or beating their quality on benchmarks. It also supports a 1M token context window, up from 260K in Mercury 2.5, and is priced at $0.20 per million input tokens and $0.70 per million output tokens. _Developers comparing fast LLM options for latency-sensitive apps track releases like this on daily.dev._ ### How much bigger is Mercury 3 than the original Mercury diffusion model and what does that mean for scaling diffusion LLMs? Mercury 3 has more than 40x the parameters of Mercury 1 (mini), and intelligence on standard benchmarks rose roughly 3x between the two versions. This is presented as evidence that the scaling laws known from autoregressive models also hold for diffusion-based language models, answering an open question about whether diffusion LLMs could compete with autoregressive models on complex tasks. _Anyone weighing diffusion versus autoregressive architectures can follow scaling results like these on daily.dev._ ### What is the pricing for Inception's Mercury 3 API and is there a launch discount? Mercury 3 is priced at $0.20 per million input tokens and $0.70 per million output tokens, but at launch it is 50% off, making it $0.10 per million input and $0.35 per million output tokens. Cached tokens receive an additional 90% discount off the input price. It is available via the Inception API Platform, OpenRouter, Baseten, and Vercel, and is OpenAI API compatible. _Teams budgeting for LLM API costs can keep tabs on pricing shifts like this via daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 82 comments across hackernews (as of 2026-10-20). **TL;DR:** Early testers are impressed by Mercury 3's speed and surprisingly solid coding/agent performance at a low price, though real-world token-per-second results and time-to-first-token sometimes fall short of the advertised numbers, and several people are frustrated the weights and model size remain closed. **Sentiment:** 45% positive · 40% mixed · 15% skeptical **The case for** - Several testers report genuinely fast, cheap, and pleasantly capable output for coding subagents, UML diagrams, and pelican-SVG style tests. - The 1M context window at low price compares very favorably to competitors like Haiku 5.5 once long-context pricing is factored in. - Low latency is seen as a real differentiator for user-facing or agentic use cases like browser agents and live-data processing. **The pushback** - Actual measured throughput (e.g. ~160 TPS) fell well short of the claimed 1,000+ TPS for at least one user. - Time-to-first-token and latency were called disappointing or slow by some, undercutting the core speed pitch. - No open weights or disclosed model size, which several commenters found disappointing. - Coding/agentic capability is seen as behind Opus/Fable-tier frontier models, more suited to subagent tasks. **By community** - hackernews (mixed): Generally enthusiastic early impressions on speed and coding ability, tempered by requests for more transparency (weights, model size, benchmark methodology) and some reports of underwhelming real-world speed/TTFT. **Hottest debate:** Whether diffusion LLMs like Mercury 3 actually deliver on their speed promise at scale, given conflicting real-world throughput reports versus the advertised benchmarks, echoing why Google's own diffusion efforts reportedly stalled at scale. **Open questions** - What is Mercury 3's actual parameter count? - Will Inception ever release open weights? - Why do real-world throughput and time-to-first-token sometimes fall well short of the advertised 1,000+ TPS? - How well does diffusion generation scale with batch size compared to autoregressive models in production? **Highlights** > Any idea why you only got 160TPS when the blog post claims 1000+TPS? That seems like the only draw to this model right now while capabilities are behind the frontier. > — [bigmadshoe on hackernews](https://news.ycombinator.com/item?id=50158434) > This talk serves as sort on update on where they are with it: https://youtu.be/r305-aQTaU0?si=j3IqG9BAT0uumPeA At a high level: It doesn't benefit from batching to reduce memory bandwidth in the same way that a "standard" autoregressive model does, which ends up meaning that it's great at reducing latency for individual/low-concurrency users and delivers better outcomes on certain tasks but much more expensive per token at scale. > — [clhodapp on hackernews](https://news.ycombinator.com/item?id=50157009) > Difussion Gemma performed a lot worse than its equivalent autoregressive counterpart (it was particularly bad because rather than a sort of constant-ish gap you might see between 2 model classes, the gap seemed to worsen the harder the task/benchmark) and the speed benefits got negated at scale. Not really surprising we haven't heard much. > — [famouswaffles on hackernews](https://news.ycombinator.com/item?id=50156658) > Quick review: It's really fast and works really well in my initial tests. I used it to build a little Swift UI app that runs in the background and helps me OCR and translate text[1]. When I hit CMD + Shift + O it selects a portion of the screen (think of the screenshot tool), runs OCR on the portion and throws that at the google translate API and displays the result in a tooltip above / below the portion. The hotkey / key combination, translation provider and target language can be changed in the settings and there is a little tray icon from which I can access those and enable/disable the tool. 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. The code has tests and looks okay for what it is and it works really well so far. All of it took me around than 45 minutes of prompting, testing, correcting and approving changes. It used 26 million tokens (1.3M uncached) at around 160TPS. I used Oh My Pi as a harness. [1]When I study chinese or japanese the PDFs I have often don't work with the macos preview selection tool and I don't want to go paste whatever I can select into another tool so I find this useful. > — [Escapado on hackernews · 2 comments](https://news.ycombinator.com/item?id=50157244) > Google more or less ordinated the category. They have now seemingly dropped it, ostensibly because it doesn’t scale economically like transformers do (via batching requests). I think though, that if Inception can keep up these speeds and prices, assuming diffusion continues to improve (likely), that Flash and Luna could lose out. > — [willy\_k on hackernews](https://news.ycombinator.com/item?id=50155941) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50155245) · 185 points · 82 comments ## Similar posts on daily.dev - [Inception Labs says its diffusion LLM is 10x faster than Claude, ChatGPT, Gemini](https://daily.dev/posts/inception-labs-says-its-diffusion-llm-is-10x-faster-than-claude-chatgpt-gemini-x9mtyj6ul) · The New Stack · 1 upvotes · 0 comments - [Inception Labs: Making LLMs Faster and More Cost-Efficient](https://daily.dev/posts/inception-labs-making-llms-faster-and-more-cost-efficient-jwseakaal) · The New Stack · 1 upvotes · 0 comments - [Inception raises $50 million to build diffusion models for code and text](https://daily.dev/posts/inception-raises-50-million-to-build-diffusion-models-for-code-and-text-fnvf3qtgf) · TechCrunch · 0 upvotes · 0 comments - [Gemini 3.1 Flash-Lite: Built for intelligence at scale](https://daily.dev/posts/gemini-3-1-flash-lite-built-for-intelligence-at-scale-oz1mpgrn8) · DeepMind · 1 upvotes · 1 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents) [View this post on daily.dev](https://daily.dev/posts/mercury-3-diffusion-at-scale-lnpntxxhe) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Mercury 3: Diffusion at scale","url":"https://daily.dev/posts/mercury-3-diffusion-at-scale-lnpntxxhe","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/mercury-3-diffusion-at-scale-lnpntxxhe"},"datePublished":"2026-10-19T18:21:35.093Z","dateModified":"2026-10-20T15:55:49.142Z","description":"Inception Labs launched Mercury 3, a diffusion-based LLM it claims is the most capable released to date, outperforming GPT-6 Luna and Claude Haiku 5.5 while...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8fcb6bdd6fab757f2f163214fe57a63e?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8fcb6bdd6fab757f2f163214fe57a63e?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/mercury-3-diffusion-at-scale-lnpntxxhe","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"Mercury 3: Diffusion at scale"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/mercury-3-diffusion-at-scale-lnpntxxhe#faq","mainEntity":[{"@type":"Question","name":"How fast is Inception's Mercury 3 diffusion model compared to GPT-6 Luna and Claude Haiku 5.5?","acceptedAnswer":{"@type":"Answer","text":"Mercury 3 runs 4-8x faster than GPT-6 Luna and Claude Haiku 5.5, generating over 1,000 tokens per second (1,068 tokens per second on NVIDIA GPUs) while matching or beating their quality on benchmarks. It also supports a 1M token context window, up from 260K in Mercury 2.5, and is priced at $0.20 per million input tokens and $0.70 per million output tokens. Developers comparing fast LLM options for latency-sensitive apps track releases like this on daily.dev."}},{"@type":"Question","name":"How much bigger is Mercury 3 than the original Mercury diffusion model and what does that mean for scaling diffusion LLMs?","acceptedAnswer":{"@type":"Answer","text":"Mercury 3 has more than 40x the parameters of Mercury 1 (mini), and intelligence on standard benchmarks rose roughly 3x between the two versions. This is presented as evidence that the scaling laws known from autoregressive models also hold for diffusion-based language models, answering an open question about whether diffusion LLMs could compete with autoregressive models on complex tasks. Anyone weighing diffusion versus autoregressive architectures can follow scaling results like these on daily.dev."}},{"@type":"Question","name":"What is the pricing for Inception's Mercury 3 API and is there a launch discount?","acceptedAnswer":{"@type":"Answer","text":"Mercury 3 is priced at $0.20 per million input tokens and $0.70 per million output tokens, but at launch it is 50% off, making it $0.10 per million input and $0.35 per million output tokens. Cached tokens receive an additional 90% discount off the input price. It is available via the Inception API Platform, OpenRouter, Baseten, and Vercel, and is OpenAI API compatible. Teams budgeting for LLM API costs can keep tabs on pricing shifts like this via daily.dev."}}]} ```