Mercury 3 Diffusion LLM: 1,068 Tokens/Sec [2026]

Inception Labs introduced Mercury 3 on October 19, 2026, a diffusion based large language model the company says generates text at 1,068 tokens per second on Nvidia GPUs. The release matters less for its raw speed number than for what it represents: a serious, well funded bet that diffusion architectures, long associated with image generation tools, can replace the token by token transformer approach that has powered every major chatbot since GPT-3.

Mercury 3 arrives in the middle of one of the most crowded model release windows of the year. Google shipped Gemini 4 Argon on October 13. Anthropic cut prices on Claude Haiku 5.5 on October 7. Microsoft AI pushed out MAI-Code-1.1-Flash the same week. Mistral announced Mistral Large 4 on October 18. Against that backdrop, Inception is not trying to out reason its rivals. It is trying to out run them, and it is doing so with a company built from the start around a different way of generating text.

Google · Preferred Sources

Don't miss new tech stories on Google

Add TrendinTech once in the Google app and our stories appear in your news suggestions.

Add Now

What Inception Labs Actually Announced on October 19

According to Inception’s own launch post, Mercury 3 is a general purpose diffusion large language model, what the company calls a dLLM, positioned as the successor to Mercury 2.5. The company says the new model expands context handling, lifts raw throughput past 1,000 tokens a second, and cuts standard pricing compared with earlier releases. Inception’s announcement frames Mercury 3 as proof that diffusion based generation scales the same way transformer stacks have scaled for the past several years, just along a different axis: parallel refinement instead of sequential prediction.

Unite.AI, which covered the launch in detail, reported that Mercury 3 is being pitched directly at latency sensitive products, voice assistants, coding copilots, and real time tutoring tools, rather than at the open ended reasoning tasks where transformer heavyweights still dominate. That framing is deliberate. Inception is not claiming Mercury 3 beats every model on every task. It is claiming Mercury 3 answers faster, which for a narrow but growing set of products matters more than a marginal reasoning gain. Unite.AI’s analysis lays out the company’s early access production results alongside the broader Mercury model family.

The timing is notable too. Mercury 3 lands one day after Mistral’s Large 4 announcement and twelve days after Anthropic’s Claude Haiku 5.5 price cut, discussed in detail in our coverage of the Claude Haiku pricing war. Every major lab is racing to ship something new before the fourth quarter closes, and speed and price have become the two levers every vendor is pulling at once.

How Diffusion Language Models Differ From Transformers

Most production chatbots, including the GPT, Claude, and Gemini families, use autoregressive transformers. They generate text one token at a time, with each new token conditioned on every token that came before it. That sequential dependency is part of why longer answers take longer to produce: the model cannot predict token 500 until it has already produced token 499.

Sequential Generation in Transformers

Autoregressive models are trained to predict the next token given everything before it. This produces coherent, high quality text, but generation speed is tightly coupled to hardware parallelism tricks like speculative decoding, since the core loop is inherently one token after another.

Parallel Refinement in Diffusion Models

Diffusion language models start from a masked or corrupted version of an output and refine multiple token positions at once across several denoising steps, a method borrowed from image generation diffusion models like those behind popular AI art tools. Because many tokens get updated in the same pass rather than one at a time, throughput can scale differently than it does for transformers. Inception’s own account of Mercury 3’s generation speed, a reported 1,068 tokens per second on Nvidia GPUs, is the company’s central piece of evidence that this approach pays off in practice.

It is worth being precise about what is confirmed and what is not. Inception has published the speed figure and the broad architectural description. It has not published a full numeric benchmark table showing exactly how far ahead or behind Mercury 3 sits relative to rival models on standard evaluation suites. That gap between a strong speed claim and a thin quality comparison is the central tension running through this entire launch.

The Numbers Behind the 1,068 Tokens a Second Claim

Inception says Mercury 3 exceeds 1,000 tokens per second and runs four to eight times faster than the speed optimized models it was benchmarked against, according to the company’s own figures reported by Startup Fortune. That is a wide range, four to eight times, which suggests the exact multiple depends heavily on which rival model and which hardware configuration is used for the comparison. Startup Fortune’s report on the launch frames the 1,068 tokens per second figure as Inception’s headline number, run on Nvidia GPUs under conditions the company has not fully disclosed.

For context, that speed matters most in applications where a human is waiting on the other end of the exchange in real time. A five second pause before an AI model starts replying is barely noticeable in a chat window where the user is reading a long answer anyway. It is a much bigger problem in a voice interface, a pair of smart glasses, or a live tutoring session where the entire interaction model depends on instant back and forth. That is precisely the market Inception is chasing with this release, rather than the long form reasoning and coding benchmarks where OpenAI, Anthropic, and Google have concentrated their marketing.

It is also worth comparing this throughput framing against what Liquid AI did with its on device model line, where the pitch was never about beating cloud models on raw quality but about running usefully fast on local hardware. Our earlier look at Liquid AI’s 24B model hitting 184 tokens a second on device covers a similar bet: that for a meaningful slice of real products, speed and efficiency beat raw benchmark supremacy.

Context Window Jumps to 1 Million Tokens

Mercury 3 expands its context window to 1,000,000 tokens, up from the 260,000 token window Inception shipped with Mercury 2.5. That is close to a fourfold increase in a single generation, and it puts Mercury 3’s context handling in the same rough bracket as the largest windows offered by rival frontier labs this year.

A separate technical analysis cited a maximum output figure of 131,072 tokens for Mercury 3, though that number did not appear directly in Inception’s own announcement excerpt, so it should be treated as a secondary, less certain data point rather than an officially confirmed specification. What is confirmed directly from Inception is the headline jump from 260,000 to 1,000,000 tokens of context, a change the company is positioning as critical for processing long documents, large codebases, and extended conversation histories without truncation.

A larger context window on its own does not say much about reasoning quality over that window, a distinction that has tripped up several model releases this year. Gemini 4 Argon’s own context and reasoning claims drew scrutiny before launch, something we covered when the UK’s AI Security Institute flagged risks around Gemini 4 Argon ahead of its release. The lesson carries over here: a published context window number is a ceiling, not a guarantee of usable performance at that ceiling.

Mercury 3 Pricing: Launch Discounts and Cached Tokens

Inception set standard pricing for Mercury 3 at $0.20 per million input tokens and $0.70 per million output tokens. For launch, the company is offering a 50 percent discount on both, bringing input down to $0.10 per million tokens and output down to $0.35 per million tokens. Cached input tokens get an additional 90 percent discount off the standard input price, a mechanism increasingly common across the industry for repeated or templated prompts.

Pricing Tier	Input (per 1M tokens)	Output (per 1M tokens)	Notes
Standard rate	$0.20	$0.70	Published baseline pricing from Inception
Launch discount	$0.10	$0.35	50 percent off standard, time limited promotion
Cached input tokens	Up to 90 percent off input price	Not applicable	Applies to repeated or templated prompt segments
Mercury 2.5 context window	260,000 tokens	N/A	Previous generation, for comparison
Mercury 3 context window	1,000,000 tokens	N/A	Confirmed in Inception’s launch announcement

These numbers slot Mercury 3 into a pricing war that has been building across the entire AI model industry since the start of the fourth quarter. Anthropic’s 90 percent price cut on Claude Haiku 5.5, which we covered in our report on the AI pricing war sparked by Claude Haiku 5.5, set a tone that every subsequent release this quarter has had to answer. Mistral’s own Large 4 launch leaned hard on cost per token in its pitch against OpenAI’s GPT-6 Astra, a comparison we broke down in our piece on Mistral Large 4 versus GPT-6 Astra pricing. Mercury 3’s launch discount fits the same pattern: undercut on price while a new architecture is still being evaluated by the market.

Benchmark Claims Against GPT-6 Luna and Claude Haiku 5.5

Inception says Mercury 3 outperforms OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 across a broad range of standard benchmarks. A launch report identified three named benchmark suites used in that comparison: IFBench, GPQA Diamond, and SWE-Bench Atlas. None of those reports, including Inception’s own materials, published the exact numeric scores Mercury 3 achieved against either rival model on those suites.

That is a meaningful gap. A speed claim backed by a specific, verifiable number, 1,068 tokens a second, sits next to a quality claim that is comparative rather than quantitative. Readers are told Mercury 3 wins, but not by how much, and not against which exact prompt sets or scoring methodology. OrcaRouter’s independent writeup of the launch made this exact point directly, describing the quality claim as comparative rather than quantitative, a framing that is hard to argue with given what Inception has actually published so far.

No comparison against any Gemini model appeared in the launch materials at all, which is itself worth noting given that Google’s Gemini 4 Argon has set the pace for reasoning and coding benchmarks since its release on October 13, a model we compared directly against Anthropic’s own flagship in our analysis of Gemini 4 Argon versus Claude Opus 5.5 on coding tasks. Leaving Gemini out of the comparison set entirely could mean Inception simply has not run that test yet, or that the result was not favorable enough to publish. Both are plausible, and neither is confirmed.

Where Mercury 3 Is Available Today

OpenRouter made Mercury 3 available on launch day, according to reporting on the release, giving developers an immediate way to test the model against alternatives without integrating directly against Inception’s own endpoint first. That kind of same day aggregator listing has become close to standard practice for major model launches this year, letting independent developers run their own side by side comparisons well before any lab publishes a full benchmark table.

Beyond the OpenRouter listing, the available reporting does not spell out every detail of Mercury 3’s software developer kit or the complete terms of its API access. What is established is the pricing structure already covered above and the fact that the model is reachable through a standard API rather than being limited to a closed pilot program. RuntimeWire’s coverage of the launch, which anchors its own reporting on the 1,068 tokens per second figure, treats same day API and aggregator availability as central to how Inception is positioning this release against slower, more gated rollouts from larger labs.

Stefano Ermon on Scaling Diffusion Models

Stefano Ermon, co-founder of Inception, addressed the central skepticism around diffusion language models directly in the company’s launch materials. “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,” Ermon said, according to Inception’s official announcement.

That parameter comparison, more than 40 times the size of the original Mercury model, is one of the more concrete figures Inception has shared, even without an absolute parameter count attached to it. It frames Mercury 3 less as an incremental update and more as a proof point that the company’s founding architectural bet holds up as the model gets bigger, which has historically been the hardest test for any new model family to pass.

Early users building on Mercury 3 echoed the latency focused pitch. Michael Tong, co-founder and chief technology officer at General Learning, the company behind RevisionDojo, described the model’s value for tutoring applications as delivering rapid feedback “without giving up reasoning,” per comments reported by Unite.AI. Kuan Zou, founder and chief executive of Octen AI, framed the same tradeoff in terms of hardware constraints, noting that for AI responses delivered through smart glasses, “an answer that takes five seconds feels broken,” also per Unite.AI’s reporting. Both comments point at the same underlying argument Inception is making: that raw benchmark leadership matters less than response time for an entire category of products that simply cannot tolerate multi second pauses.

Competitive Comparison: Mercury 3 Versus the October 2026 Model Wave

Mercury 3 did not launch into an empty field. October 2026 has been one of the busiest single months for frontier and near frontier model releases this year, with at least seven named models shipping or being announced inside a three week window.

Model	Developer	Category	October 2026 Date
Mercury 3	Inception Labs	Diffusion LLM (dLLM)	October 19
Mistral Large 4	Mistral	General purpose, multimodal, reasoning	Announced October 18, public October 27
Gemini 4 Argon	Google	General, reasoning, coding	October 13
Claude Haiku 5.5	Anthropic	General, reasoning, coding	October 7
MAI-Code-1.1-Flash	Microsoft AI	Coding	October 7
Nano Banana 2.1 / EmbeddingGemma 2	Google DeepMind	Multimodal / embedding	October 6
Beam	Reflection AI	General, reasoning, coding	October 5

Against that table, Mercury 3 stands out as the only diffusion based entry. Every other named model on the list still uses some variant of the autoregressive transformer approach. That makes Mercury 3 less a direct competitor to any single model on speed and reasoning both, and more a test case for whether a structurally different approach can carve out a durable niche alongside the dominant architecture, the same way OrcaRouter framed its own comparison of Mercury 3’s clock speed against DeepSeek V4 Pro’s launch pricing, treating the two models as answering different questions rather than competing head to head on the same axis.

DeepSeek’s own release cadence this year, including the V4.1-Flash-1017 model that opened its search API, shows how quickly the lower cost end of the model market has been moving in parallel with the frontier labs. Inception’s bet is that diffusion gives it a structural speed advantage that undercutting on price alone cannot replicate, a different kind of competitive moat than the one DeepSeek or Mistral are building.

Historical Context: From Mercury 1 to Mercury 3

Mercury 1 and Mercury 2.5

Inception first introduced the Mercury line as a general chat diffusion large language model, staking out diffusion based text generation as the company’s core differentiator well before most of the industry took the approach seriously. Mercury 2.5 followed with a 260,000 token context window, a respectable figure at the time but far short of what frontier transformer models were already offering.

The Jump to Mercury 3

Mercury 3 represents the point at which Inception is asking the market to take diffusion seriously as a frontier approach rather than a research curiosity. A context window near four times larger, a parameter count more than 40 times that of the original Mercury model, and a speed figure north of 1,000 tokens a second together form the company’s case that this architecture has crossed from promising experiment into something enterprises can actually build on. Whether that case holds will depend heavily on independent benchmark verification that has not yet happened.

Market Impact: What Faster Inference Means for Developers

For developers building latency sensitive products, a credible 1,068 tokens per second figure changes what is technically possible without necessarily changing anything else about how a team writes its prompts or structures its application. Voice agents, live captioning tools, and real time coding assistants have all historically had to engineer around model latency with tricks like speculative decoding, partial response streaming, and aggressive caching. A model that is simply faster at the source removes some of that engineering burden, assuming the quality holds up under real workloads rather than just launch day marketing claims.

The pricing war angle matters just as much for budget conscious teams. A launch discount bringing output tokens down to $0.35 per million is competitive territory in a market where GPT-6.1 Sol, Sonnet 5.5, and Gemini 3.8 Flash have been locked in their own three way price comparison, detailed in our earlier report on GPT-6.1 Sol versus Sonnet 5.5 versus Gemini 3.8 Flash pricing. Every new entrant at this price point adds pressure on the established players to either match the discount or justify a premium with demonstrable quality gains, gains Inception has so far only described in general terms rather than published scores.

Real World Use Cases: Tutoring Bots and Wearable Devices

The two named early adopters disclosed around the launch, General Learning’s RevisionDojo tutoring product and Octen AI’s smart glasses software, point at exactly the kind of product where diffusion’s parallel generation approach could matter most. A tutoring tool needs to respond to a student’s answer quickly enough to feel like a conversation rather than a query. A wearable device strapped to someone’s face has essentially zero tolerance for the kind of multi second pause that a chat window user might not even notice.

Neither use case requires Mercury 3 to beat GPT-6 Luna or Claude Haiku 5.5 on every reasoning benchmark. They require it to be fast enough, reliable enough, and cheap enough at the volumes those products generate. That is a narrower bar than the one frontier labs are competing to clear, and it may be the more realistic opening for a model family built around a genuinely different architecture to establish itself before attempting to compete on reasoning quality across the board.

Risks and Open Questions

The biggest open question is simple: how good is Mercury 3 actually, measured against a transparent, reproducible benchmark table. Until independent evaluators publish their own numbers against IFBench, GPQA Diamond, SWE-Bench Atlas, or any other widely used suite, Inception’s claims remain a comparative statement from the company that built the model, not a verified result.

A second question is durability of the pricing. Launch discounts of 50 percent are common across this entire model release cycle and tend to shrink or disappear once initial adoption numbers come in. Teams building production systems on Mercury 3’s discounted rate should expect the standard $0.20 and $0.70 per million token prices to become the real long term cost basis.

A third is competitive response. If Mercury 3’s speed claims hold up under independent testing, it would not be surprising to see transformer focused labs invest more heavily in their own parallel decoding or speculative generation techniques to close the gap, rather than conceding the latency sensitive product category to a diffusion based challenger outright.

Five Predictions for Diffusion Language Models Heading Into 2027

Independent benchmark sites will publish head to head numeric comparisons between Mercury 3, GPT-6 Luna, and Claude Haiku 5.5 within weeks, filling the gap left by Inception’s comparative rather than quantitative launch claims.
At least one more major lab will announce or expand a diffusion based model line in the next two quarters, following Inception’s lead in targeting latency sensitive product categories rather than general purpose reasoning.
Mercury 3’s launch discount pricing will not survive past the first quarter of 2027 in its current form, following the same pattern seen across other recent model launch promotions this year.
Wearable and voice first hardware makers will become the fastest growing customer segment for diffusion based inference providers, given how directly latency maps to user experience in those product categories.
Aggregator platforms like OpenRouter will keep playing a larger role in how fast new architectures reach real developer testing, often beating a model’s own native software developer kit to market by days.

Frequently Asked Questions About Mercury 3

What is Mercury 3?

Mercury 3 is a diffusion based large language model released by Inception Labs on October 19, 2026. The company describes it as a general purpose dLLM that generates text through parallel refinement rather than the token by token prediction used by transformer models.

How is a diffusion LLM different from a transformer model?

Transformer models generate text one token at a time, with each token depending on everything generated before it. Diffusion language models start from a masked or corrupted output and refine multiple token positions across several steps, which can allow for different throughput characteristics than strictly sequential generation.

How fast is Mercury 3 compared to other AI models?

Inception reports Mercury 3 generates 1,068 tokens per second on Nvidia GPUs, which the company says is four to eight times faster than the speed optimized rival models it compared against. These figures come from Inception’s own announcement and have not yet been independently verified against a standardized benchmark.

How much does Mercury 3 cost to use?

Standard pricing is $0.20 per million input tokens and $0.70 per million output tokens. A launch discount cuts both by 50 percent, to $0.10 and $0.35 per million tokens respectively. Cached input tokens receive up to a 90 percent discount off the standard input price.

Does Mercury 3 actually beat GPT-6 Luna and Claude Haiku 5.5?

Inception says Mercury 3 outperforms both models across benchmarks including IFBench, GPQA Diamond, and SWE-Bench Atlas, but the company has not published exact numeric scores for either comparison. The claim is currently comparative rather than independently quantified.

What is Mercury 3’s context window?

Mercury 3 has a 1,000,000 token context window, up from the 260,000 token window in Mercury 2.5. A separate technical analysis cited a maximum output figure of 131,072 tokens, though that number was not part of Inception’s own announcement.

Where can developers access Mercury 3?

Mercury 3 was made available through OpenRouter on launch day, alongside standard API access through Inception. Full software developer kit and integration details beyond pricing have not been fully published yet.

Who founded Inception Labs?

Stefano Ermon is a co-founder of Inception Labs and spoke on the record about Mercury 3’s scaling claims in the company’s launch materials, emphasizing that the model is more than 40 times larger than the original Mercury while still running at over 1,000 tokens per second.

Inception Labs Mercury 3 Hits 1,068 Tokens a Second [2026]