Inception Labs has introduced Mercury 3, a diffusion-based large language model designed to deliver high-quality AI responses at substantially higher speeds than conventional language models. The company says Mercury 3 generates 1,068 tokens per second on Nvidia GPUs and outperforms OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 across several benchmarks. Inception’s internal evaluation showed Mercury 3 leading in eight of 12 tests, although independent benchmarking had not yet verified those results at launch.
The model represents a different approach to building language AI. Instead of generating an answer one token at a time, Mercury 3 starts with an incomplete or noisy representation and progressively refines multiple parts of the output. Inception says this design allows it to combine fast responses with competitive model quality. The company is targeting applications such as autonomous AI agents, real-time search, interactive tutoring and customer service, where even small delays can affect the user experience.
Key Takeaways
- Mercury 3 is a diffusion language model: It refines text through multiple passes rather than relying exclusively on sequential token generation.
- Inception claims speeds above 1,000 tokens per second: Its reported result is 1,068 tokens per second on Nvidia GPUs.
- The model won eight of 12 internal benchmarks: Inception compared it with GPT-6 Luna and Claude Haiku 5.5.
- It supports a one-million-token context window: This allows it to process large amounts of text in a single context.
- Launch pricing is discounted: Promotional rates are $0.10 per million input tokens and $0.35 per million output tokens.
- Independent verification remains important: The company’s benchmark results and performance claims still require broader independent testing.
What Is Mercury 3?
Mercury 3 is the latest model from Inception Labs, a startup developing diffusion-based large language models. Its central proposition is that AI systems can scale in capability without losing the low latency that makes them useful in real-time applications.
Most widely used language models rely on autoregressive generation. They predict the next token based on the tokens that have already been generated, repeating this process until an answer is complete.
Diffusion language models use a different method. They progressively refine an output over several steps, potentially updating multiple tokens in parallel. The approach resembles the iterative refinement used by diffusion models for generating images, although language models must also preserve coherent grammar, meaning and reasoning.
This difference can affect how quickly a model produces a complete answer. Inception says Mercury 3 combines the parallel-generation approach with a substantially larger model than its first-generation Mercury system.
The company has not disclosed the model’s exact parameter count. It says Mercury 3 is more than 40 times larger than Mercury 1 and delivers roughly three times the intelligence on a range of benchmarks compared with Mercury 1 mini.
Mercury 3 vs GPT-6 Luna and Claude Haiku 5.5
Inception’s headline claim is that Mercury 3 can compete with established speed-focused models while generating responses considerably faster.
The company’s benchmark comparison covers 12 tests, including instruction following, coding and other reasoning-oriented tasks. Mercury 3 reportedly won eight of those tests, while the rival models performed better in several others.
| Performance measure | Mercury 3 | GPT-6 Luna | Claude Haiku 5.5 |
|---|---|---|---|
| Reported generation speed | 1,068 tokens/sec | About 130 tokens/sec | About 240 tokens/sec |
| Internal benchmark wins | 8 of 12 tests | Leads in selected tests | Leads in selected tests |
| Maximum context window | 1 million tokens | Varies by model and configuration | Varies by model and configuration |
| Independent Mercury 3 verification at launch | Not available | Not applicable | Not applicable |
Sources: Inception Labs and The Decoder. Speed figures are company-reported comparisons and may depend on hardware, configuration, workload and measurement methodology. Benchmark wins refer to Inception’s comparison, not a universal ranking.
Inception reported particularly strong results in instruction following and autonomous coding. Mercury 3 scored 76.84 on IFBench and 51.08 on SWE-Bench Atlas in the company’s evaluation.
However, it did not lead every test. Claude Haiku 5.5 performed better on the DeepSWE coding benchmark and a test involving retrieval from long texts. GPT-6 Luna also led in two tests from Artificial Analysis, according to The Decoder’s report.
These differences matter because AI performance cannot be captured by a single score. A model that is faster at producing an answer may not be the most accurate or dependable choice for every application.
Why Diffusion Models Could Make AI Faster
The traditional approach to language generation creates a dependency between successive tokens. Even when a model processes information efficiently, it must repeatedly predict what comes next.
Diffusion models attempt to reduce this bottleneck by refining several parts of an output at once. Rather than waiting for every token to be generated in sequence, the system can improve a larger portion of the answer during each refinement pass.
<box gap={2}> <row align=start gap={3}> <box background=”surface-secondary” radius=”lg” padding={3} size=”44px” align=center justify=center> <icon name=”list-ordered” size=”xl” /> </box> <box flex=”1″ gap={1}> **Traditional autoregressive generation**
The model predicts one token after another. Each step depends on the previous output, which can increase the time required for long responses.
</box>
</row> <divider color=”subtle” /> <row align=start gap={3}> <box background=”surface-secondary” radius=”lg” padding={3} size=”44px” align=center justify=center> <icon name=”layers” size=”xl” /> </box> <box flex=”1″ gap={1}> **Diffusion-based generation**
The model progressively refines an answer across multiple steps, updating many parts of the output together to reduce sequential generation delays.
</box>
</row> </box>
The advantage is especially relevant when an application repeatedly calls a model. An AI agent researching a website, for example, may need to navigate a page, interpret information, decide on its next action and repeat the process many times.
If each model response arrives faster, the entire workflow can become more responsive. However, total task completion time also depends on tools, network requests, external systems and the complexity of the task.
Where Inception Plans to Use Mercury 3
Inception is targeting applications that benefit from low latency and repeated AI interactions.
Autonomous AI agents
AI agents often perform multiple steps to complete a task. They may search for information, interact with websites, analyse results and decide what to do next.
Inception says Mercury 3 completed 50% of 50 live website-navigation tasks in a test conducted with TinyFish AI. The company described the result as better than GPT-6.1-Sol in that particular evaluation, but the limited task count means the result should not be treated as a broad measure of agent capability.
Real-time search and smart glasses
Search applications need to return information quickly enough to maintain a natural interaction. Inception’s early-access partner Octen AI reported a median generation time of 0.9 seconds with Mercury 3 and a median end-to-end response time of 2.2 seconds in its Sense API tests.
Such performance could be useful for smart glasses, voice assistants and other interfaces where users expect answers without a noticeable pause.
Education and customer support
Inception’s early-access partner General Learning said Mercury 3 helped deliver student feedback, hints and corrections in about one second.
Customer support is another target. Faster responses can make conversational systems feel more natural, but speed alone is insufficient for businesses handling sensitive or regulated requests. Accuracy, appropriate escalation and reliable task completion remain essential.
Mercury 3 Pricing and Availability
Inception has published standard prices of $0.20 per million input tokens and $0.70 per million output tokens. At launch, it is offering a 50% discount, reducing those rates to $0.10 and $0.35, respectively. The company has not specified how long the promotion will last.
| Pricing component | Standard price | Launch price |
|---|---|---|
| Input tokens per million | $0.20 | $0.10 |
| Output tokens per million | $0.70 | $0.35 |
Mercury 3 is available through Inception’s API platform and through OpenRouter, Baseten and Vercel’s AI Gateway. The company says its API is compatible with the OpenAI API format, potentially reducing the engineering work required for developers to test or migrate existing applications.
Compatibility does not guarantee that every model feature or behaviour will be identical. Developers should evaluate output quality, latency, reliability, rate limits and total costs before moving production workloads.
The Challenges Mercury 3 Still Needs to Address
Mercury 3’s claims are promising, but several questions remain open.
First, independent testing is needed to determine whether its speed advantage holds across different hardware, prompt lengths and production workloads. Tokens-per-second figures are not always directly comparable because benchmarking methods can differ.
Second, developers need to evaluate model quality for their own tasks. Mercury 3’s reported wins in instruction following and coding are counterbalanced by tests in which competing models performed better.
Third, diffusion-based language models must demonstrate reliable performance across long and complex tasks. High generation speed is valuable, but a model that requires additional corrections or repeated attempts may not deliver the same advantage in practice.
Finally, developers should compare actual operating costs. The launch discount makes Mercury 3 inexpensive on a per-token basis, but pricing after the promotion, context usage and the number of model calls can affect the economics of a deployment.
The Bigger Picture
Mercury 3 highlights a growing effort to improve AI not only by making models more capable, but also by changing how they generate answers. If diffusion-based language models can maintain strong reasoning and instruction-following performance while generating text substantially faster, they could be attractive for applications that depend on immediate feedback and repeated model interactions.
The announcement also reflects intensifying competition in the market for fast, affordable AI. Inception is challenging larger providers by focusing on an architectural difference rather than competing only through conventional model scaling. Whether that strategy becomes commercially significant will depend on independent benchmarks, developer adoption and real-world reliability.
FAQs
1. What is Inception’s Mercury 3?
Mercury 3 is a diffusion-based large language model developed by Inception Labs. It generates text through iterative refinement rather than relying entirely on sequential token generation.
2. How fast is Mercury 3?
Inception reports a generation speed of 1,068 tokens per second on Nvidia GPUs. This is a company-reported figure and should be compared with other models only when hardware and benchmarking conditions are similar.
3. Does Mercury 3 outperform GPT-6 Luna and Claude Haiku 5.5?
Inception says Mercury 3 won eight of 12 internal benchmark tests against the two models. The rivals led in some categories, and independent testing of Mercury 3 was not available at launch.
4. How much does Mercury 3 cost?
Its listed standard pricing is $0.20 per million input tokens and $0.70 per million output tokens. Promotional launch pricing cuts those rates to $0.10 and $0.35, respectively.
5. Who should consider using Mercury 3?
Developers building real-time AI agents, search tools, tutoring systems, voice interfaces and customer support applications may find its low-latency approach useful. They should benchmark it on their own workloads before deploying it at scale.
Looking Ahead
The next important milestone for Mercury 3 will be independent evaluation. External tests can establish whether its reported speed advantage translates into better end-to-end performance and whether its quality holds up across coding, reasoning and enterprise tasks.
For the wider AI industry, Mercury 3 is another test of whether diffusion can become a practical alternative to conventional language-model generation. If the approach continues to scale while preserving quality, developers could gain another option for building responsive AI applications without relying exclusively on established autoregressive architectures.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
