--- title: Mercury 3 says diffusion LLMs can scale, now prove it description: Inception Labs released Mercury 3, a diffusion-based language model generating over 1,000 tokens per second, about 8 times faster than GPT-6 Luna and 4.5 times... canonical: https://daily.dev/posts/mercury-3-says-diffusion-llms-can-scale-now-prove-it-5wit1h7yt twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Mercury 3 says diffusion LLMs can scale, now prove it | daily.dev og:description: Inception Labs released Mercury 3, a diffusion-based language model generating over 1,000 tokens per second, about 8 times faster than GPT-6 Luna and 4.5 times... og:url: https://daily.dev/posts/mercury-3-says-diffusion-llms-can-scale-now-prove-it-5wit1h7yt og:image: https://api.daily.dev/og/posts/5WIt1H7Yt.png og:image:alt: Mercury 3 says diffusion LLMs can scale, now prove it 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 says diffusion LLMs can scale, now prove it **[Trends](https://daily.dev/sources/trends)** · 3 min read · 0 upvotes · 0 comments ## Summary Inception Labs released Mercury 3, a diffusion-based language model generating over 1,000 tokens per second, about 8 times faster than GPT-6 Luna and 4.5 times faster than Claude Haiku 5.5. Its context window grew to 1M tokens, up from 260K, and the model is over 40 times larger than Mercury 1. Pricing differs across sources: one independent test cites $0.10/$0.35 per million tokens, while Inception's own announcement lists $0.20/$0.70 with 50% off at launch. A companion model, Mercury Decide, scored 87% on legal bench versus 74% for Jev at roughly a quarter of the cost, resolving a judgment in about 0.16 seconds. Mercury 3 is available on the Inception API, OpenRouter, Baseten, and Vercel, and is OpenAI API compatible. ## Content Diffusion language models have spent years as the "cool but not serious" corner of AI. Inception's Mercury 3 is a loud attempt to change that, and the first question people are asking is whether it's really a big model now. @scaling01 did the napkin math: Mercury 3 is over 40x the size of Mercury 1, so "if Mercury 1 was something like 7B, then Mercury 3 would be > 280B." His opening line says it all: "are diffusion LLMs taking off?" ## The claim Inception says Mercury 3 beats GPT-6 Luna, Claude Haiku 5.5 and GLM-5.3 Flash while running four to eight times faster. That works out to over 1,000 tokens per second on Nvidia chips, about 8x GPT-6 Luna and 4.5x Haiku 5.5. It reportedly matches them on benchmarks like GPQA Diamond. The context window grew to 1M tokens, up from 260K. The catch is that those numbers are self-reported, and outside testing is still thin. The Mercury family is a speed pitch first, and Inception is selling that. ## The price Launch pricing is half off the standard rate. One source puts it at $0.10 in and $0.35 out per million tokens. The company's own launch page lists $0.20/$0.70 as the standard rate, which makes the discount 50%. It's available through Inception's API, OpenRouter, Baseten and Vercel's AI Gateway, and it speaks OpenAI's API format, so trying it means swapping a base URL. ## The side bet Mercury Decide is the sleeper. It's a fast decision model that scored 87% on LegalBench versus 74% for Jev, at roughly a quarter of the cost, and it resolves a single judgment in about 0.16 seconds. Inception also launched Mercury Voice, which makes the pitch clear: real-time AI where latency is the product. Early-access partners fit that story. TinyFish AI uses it for browser agents, RevisionDojo for tutoring, Octen AI for search, and Lorikeet for customer support. Every one of those cares more about a snappy response than a slightly smarter one. ## The weird part Inception says it ran swarms of Mercury 3 agents to improve the model itself, which added about 16% speed. Treat it as a company claim until someone reproduces it. My read: if the speed holds up outside Inception's own benchmarks, a 1,000 tok/s model that matches small frontier models at these prices is a real problem for the autoregressive crowd. If it doesn't, this is a very fast way to be average. ## Questions this post answers ### How fast is Inception Labs' Mercury 3 diffusion model compared to GPT-6 Luna and Claude Haiku 5.5? Mercury 3 generates over 1,000 tokens per second, which Inception Labs describes as roughly 8 times faster than GPT-6 Luna and 4.5 times faster than Claude Haiku 5.5, with similar performance on intelligence benchmarks like GPQA Diamond. Its context window also expanded to 1 million tokens, up from 260K in the prior version. _Teams rewriting agent loops for latency can track diffusion LLM benchmarks like this on daily.dev._ ### What is the pricing for Mercury 3 from Inception Labs? Pricing is inconsistent across sources: one independent test lists $0.10 per million input tokens and $0.35 per million output tokens, while Inception's own announcement states $0.20/$0.70 with 50% off at launch. The two figures are only consistent if the lower number reflects the discounted launch price, so checking the official pricing page before budgeting is recommended. _Anyone budgeting around a new model's launch pricing can follow clarifications like this on daily.dev._ ### How does Mercury Decide compare to Jev on legal benchmark tasks? Mercury Decide, a fast decision model from Inception Labs, scored 87% on legal bench compared to 74% for Jev, at about a quarter of the cost and resolving a single judgment in roughly 0.16 seconds. It's part of a broader real-time AI portfolio that also includes Mercury Voice. _Developers comparing decision models on cost and speed can follow benchmarks like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 125 comments across hackernews, x (as of 2026-10-20). **TL;DR:** Reactions focus on whether Mercury 3's claimed 40x size increase and 1000+ tok/s speed actually hold up, with a lot of skepticism about the unverified 280B parameter estimate and benchmark comparison choices. Mercury 3: Diffusion at scale: 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:** 41% positive · 42% mixed · 17% skeptical **The case for** - Parallel decoding at this scale is seen as a genuinely interesting advance if it holds up, since speed usually degrades as models get bigger. - Several testers report genuinely fast, cheap, and pleasantly capable output for coding subagents, UML diagrams, and pelican-SVG style tests. - Some see clear value for agentic workloads where per-token latency cost matters more than raw capability. - The 1M context window at low price compares very favorably to competitors like Haiku 5.5 once long-context pricing is factored in. - A few note that fast, cheap generation is exactly what routine agent/tool-calling steps need. **The pushback** - Several argue comparing against 'medium effort' Luna/Haiku configurations is misleading since few serious users run that tier. - Actual measured throughput (e.g. ~160 TPS) fell well short of the claimed 1,000+ TPS for at least one user. - The 280B parameter figure is repeatedly called a guess derived from assumptions about Mercury 1's size, not a disclosed number. - Time-to-first-token and latency were called disappointing or slow by some, undercutting the core speed pitch. - Pricing is inconsistent between the discounted launch price and Inception's own listed price, undercutting the cost-advantage claim. **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. - x (mixed): Replies swing between genuine excitement about diffusion models scaling to frontier size and pointed skepticism about unverified parameter counts, pricing claims, and benchmark framing. **Hottest debate:** Whether comparing Mercury 3 to 'medium effort' Luna/Haiku is a fair transparency move or a misleading benchmark choice. **Open questions** - What is the actual parameter count and training compute behind Mercury 3? - What is Mercury 3's actual parameter count? - Does the 1000+ tok/s speed advantage survive at larger scale and in long-context use? - Will Inception ever release open weights? - How does schema/output precision hold up under the diffusion denoising process? **Highlights** > @scaling01 Comparing to medium effort is highly sus Nobody seriously uses haiku/luna med > — [\_Teskh on x · 28 points, 4 comments](https://x.com/_Teskh/status/2112240061379162126) > @_Teskh @scaling01 luna med is the default for a billion chatgpt users. > — [Ark\_\_PL on x · 21 points](https://x.com/Ark__PL/status/2112242546810474509) > @scaling01 $0.10 / $0.35 in that table is the 50% launch price. Inception's blog lists Mercury 3 at $0.20 / $0.70. At list, that's 2x Luna and Haiku on input and 1.4x on output, using the same table. The 40x is total parameters, and the blog gives no count for either model. > — [indrajeet877 on x · 5 points](https://x.com/indrajeet877/status/2112240802659532965) > @scaling01 this would be good if Deepseek V4.1-Flash didn't have 3500tok/s 3rd party providers now at a fraction of a price per task than Mercury 3 while also having much better overall capability > — [follistatindev on x · 1 points](https://x.com/follistatindev/status/2112304736934441278) > @scaling01 Diffusion had one job: speed. If scaling to 280B just buys back the latency it saved, it's a treadmill. The real question isn't parameter count, it's whether tokens-per-second survives the scale-up. > — [Trulyammar on x](https://x.com/Trulyammar/status/2112465717396422876) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50155245) · 185 points · 82 comments - [x](https://x.com/scaling01/status/2112233800583569747) · 0 points · 43 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 - [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), [#openrouter](https://daily.dev/tags/openrouter) [View this post on daily.dev](https://daily.dev/posts/mercury-3-says-diffusion-llms-can-scale-now-prove-it-5wit1h7yt) ```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 says diffusion LLMs can scale, now prove it","url":"https://daily.dev/posts/mercury-3-says-diffusion-llms-can-scale-now-prove-it-5wit1h7yt","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/mercury-3-says-diffusion-llms-can-scale-now-prove-it-5wit1h7yt"},"datePublished":"2026-10-20T13:22:32.965Z","dateModified":"2026-10-20T15:56:04.683Z","description":"Inception Labs released Mercury 3, a diffusion-based language model generating over 1,000 tokens per second, about 8 times faster than GPT-6 Luna and 4.5 times...","image":"https://i.ytimg.com/vi/bEO_QeJU7U0/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/bEO_QeJU7U0/sddefault.jpg","isAccessibleForFree":true,"articleSection":"Trends","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":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/mercury-3-says-diffusion-llms-can-scale-now-prove-it-5wit1h7yt","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,openrouter","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Mercury 3 says diffusion LLMs can scale, now prove it"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/mercury-3-says-diffusion-llms-can-scale-now-prove-it-5wit1h7yt#faq","mainEntity":[{"@type":"Question","name":"How fast is Inception Labs' Mercury 3 diffusion model compared to GPT-6 Luna and Claude Haiku 5.5?","acceptedAnswer":{"@type":"Answer","text":"Mercury 3 generates over 1,000 tokens per second, which Inception Labs describes as roughly 8 times faster than GPT-6 Luna and 4.5 times faster than Claude Haiku 5.5, with similar performance on intelligence benchmarks like GPQA Diamond. Its context window also expanded to 1 million tokens, up from 260K in the prior version. Teams rewriting agent loops for latency can track diffusion LLM benchmarks like this on daily.dev."}},{"@type":"Question","name":"What is the pricing for Mercury 3 from Inception Labs?","acceptedAnswer":{"@type":"Answer","text":"Pricing is inconsistent across sources: one independent test lists $0.10 per million input tokens and $0.35 per million output tokens, while Inception's own announcement states $0.20/$0.70 with 50% off at launch. The two figures are only consistent if the lower number reflects the discounted launch price, so checking the official pricing page before budgeting is recommended. Anyone budgeting around a new model's launch pricing can follow clarifications like this on daily.dev."}},{"@type":"Question","name":"How does Mercury Decide compare to Jev on legal benchmark tasks?","acceptedAnswer":{"@type":"Answer","text":"Mercury Decide, a fast decision model from Inception Labs, scored 87% on legal bench compared to 74% for Jev, at about a quarter of the cost and resolving a single judgment in roughly 0.16 seconds. It's part of a broader real-time AI portfolio that also includes Mercury Voice. Developers comparing decision models on cost and speed can follow benchmarks like this on daily.dev."}}]} ```