Mercury 2.5 Brings Diffusion AI to Faster Coding, Search and Voice Agents

Inception is taking a different approach with Mercury 2.5, its newest production diffusion large language model.

Announced on September 8, 2026, Mercury 2.5 is designed to provide a substantial quality improvement over Mercury 2 while maintaining the low-latency characteristics of Inception's diffusion-based architecture.

The company reports a headline speed of 1,107 tokens per second on widely available NVIDIA GPUs.

It also offers a 260K-token context window, tunable reasoning, parallel tool calls and schema-aligned JSON output.

Those features make Mercury 2.5 particularly interesting for AI agents, search systems, voice applications and coding assistants where every additional model call can add latency.

What Is Mercury 2.5?

Mercury 2.5 is Inception's latest diffusion-based large language model.

Unlike conventional autoregressive LLMs that generate text sequentially, diffusion language models use a different generation process.

Inception's approach is designed around producing and refining tokens in parallel.

The company says Mercury 2.5 is its most capable production model yet and the largest diffusion language model it has trained.

The model is intended for applications where speed matters as much as intelligence.

That includes:

  • Search agents
  • Retrieval-augmented generation
  • Coding assistants
  • Coding subagents
  • Voice agents
  • Interactive applications
  • Tool-using AI systems
  • Real-time AI experiences

The goal is not simply to make another general chatbot.

Inception is targeting workloads where an AI model may need to be called repeatedly.

The 1,107 Tokens-Per-Second Claim

The most eye-catching number from the launch is speed.

Inception reports that Mercury 2.5 can reach 1,107 tokens per second on widely available NVIDIA GPUs.

This is important because latency becomes increasingly expensive when AI systems use multiple model calls.

Consider an AI search agent.

One user request could require the system to:

  1. Understand the question.
  2. Rewrite the search query.
  3. Search multiple sources.
  4. Rerank results.
  5. Extract information.
  6. Summarize evidence.
  7. Check the answer.
  8. Generate the final response.

Each step can require another model call.

If every call takes several seconds, the overall user experience becomes slow.

A faster model can reduce the delay across the entire workflow.

Diffusion LLMs Work Differently

Traditional large language models generally generate text one token after another.

For example:

Token 1 → Token 2 → Token 3 → Token 4

The process is sequential.

Diffusion language models approach generation differently.

They can start with a rough representation and progressively refine the output.

This creates opportunities for parallel processing.

The result can be significantly lower latency when the underlying architecture and hardware are optimized for the approach.

Inception has been developing diffusion-based language models specifically around this idea.

Mercury 2.5 represents the company's latest attempt to make that architecture competitive for real production workloads.

Mercury 2.5 Is 40% More Intelligent Than Mercury 2

Inception says Mercury 2.5 provides a 40% increase in intelligence compared with Mercury 2.

The company compares the new model's capabilities with cost-optimized frontier models such as GPT-5.6 Luna Low, Gemini 3.5 Flash-Lite and Claude Haiku 4.5.

These are Inception's comparisons rather than independent benchmark results.

The important point is that Inception is attempting to improve both sides of the equation:

More capability + low latency

That combination is particularly valuable for production AI applications.

260K Context Window

Mercury 2.5 supports a 260K-token context window.

A large context window allows an application to provide significantly more information to the model in a single interaction.

This can be useful for:

  • Large codebases
  • Long documents
  • Search results
  • Extended agent sessions
  • Technical documentation
  • Conversation history
  • Tool outputs

For coding agents, context length is especially important.

A coding agent may need to understand multiple files, tool outputs, previous decisions and project requirements at the same time.

A larger context window can reduce the need to repeatedly compress or discard information.

Tunable Reasoning Comes to Mercury 2.5

Mercury 2.5 also supports tunable reasoning.

This gives developers more control over the trade-off between reasoning depth, latency and output.

Not every request requires the same amount of computation.

A simple classification task may need only a quick response.

A complex coding problem may benefit from deeper reasoning.

Tunable reasoning allows developers to adjust the model according to the workload.

This is becoming increasingly important as AI applications move from simple chat interfaces toward agentic systems.

Parallel Tool Calls Could Help AI Agents

Another notable feature is parallel tool calling.

AI agents frequently need to use external tools.

For example, an agent might need to check:

  • A database
  • A calendar
  • A web search
  • A documentation system

Some of those operations can potentially happen at the same time.

If the model can request multiple tools in parallel, an agent does not necessarily need to wait for each tool operation to complete before starting the next one.

That can reduce total workflow latency.

For developers building agentic systems, this may be more important than raw token-generation speed.

Schema-Aligned JSON for Applications

Mercury 2.5 also supports schema-aligned JSON output.

Structured output is important when AI responses need to be consumed by software rather than directly displayed to a person.

For example, an application may ask the model to return:

  • Product information
  • Search results
  • Database operations
  • Tool parameters
  • Classification results
  • Workflow instructions

If the output follows a predefined schema, software can process the response more reliably.

This makes Mercury 2.5 more suitable for production applications than a model designed primarily for conversational text.

Mercury 2.5 Is Already Being Used for AI Search

Inception says Mercury is already being used in production search infrastructure.

The company explains that one search request can trigger many model calls.

Those calls can include:

  • Query planning
  • Query rewriting
  • Result reranking
  • Fact structuring
  • Source summarization
  • Answer checking

Because Mercury is designed for low latency, Inception says it can keep these repeated calls fast enough for a single user interaction.

This is an important use case for diffusion models.

A model does not necessarily need to be the biggest or most powerful model available if it can perform a supporting task quickly.

Mercury 2.5 Is Also Built for Voice Agents

Voice applications are another area where latency matters.

When someone speaks to an AI phone agent, delays are immediately noticeable.

A response that takes several seconds can make the conversation feel unnatural.

Inception says its production partner OpenCall achieved median model response latency close to 170 milliseconds using Mercury.

The company also reports that OpenCall reduced its P99 response time from several minutes to about one second and its P50 response time from 0.4 seconds to under 0.2 seconds after switching to Mercury.

These are partner-reported production results rather than independent benchmarks.

Still, they demonstrate the type of application Inception is targeting.

Coding Agents Could Benefit From Mercury 2.5

Coding agents are another major use case.

Modern coding agents often divide work across multiple models.

One model may plan.

Another may search documentation.

Another may summarize tool output.

Another may compress context.

Another may generate code.

Every supporting call adds latency and cost.

Inception says Augment Code uses Mercury for context compaction, model routing and MCP tool search.

According to Inception, moving context compaction to Mercury reduced latency by 82%, from around 150 seconds to 27 seconds, while reducing cost by 90% without reducing quality.

Again, these are vendor-reported results from a production customer.

But they illustrate an important idea.

The fastest model does not necessarily need to write the final answer.

It can handle the repetitive supporting tasks that surround a more capable model.

Mercury 2.5 Can Generate Applications

Inception also demonstrated Mercury 2.5 generating a working music-discovery web application from a small number of prompts.

This shows how the company sees the model fitting into broader coding workflows.

Fast models can be useful for interactive development because developers want quick feedback while working.

A model that takes several seconds for every request can interrupt the development process.

A much faster model can make AI-assisted coding feel more like an interactive development tool.

Mercury Voice and Mercury Router

Inception did not stop at Mercury 2.5.

The company also announced previews of Mercury Voice and Mercury Router.

Mercury Voice

Mercury Voice is designed specifically for voice agents.

Inception says it targets time-to-first-token below 170 milliseconds and is optimized for applications with extremely tight latency requirements.

That could make it useful for conversational AI systems where users expect near-immediate responses.

Mercury Router

Mercury Router takes a different approach.

Instead of generating the final answer itself, it analyzes incoming prompts and routes them to an appropriate model.

Inception says it can route requests to both open and closed models based on the desired balance between:

  • Quality
  • Speed
  • Cost

This is an interesting development because AI applications increasingly use multiple models rather than relying on one model for everything.

Mercury Router Could Reduce AI Infrastructure Costs

AI developers often face a difficult question:

Which model should handle each request?

Using a powerful frontier model for every task can become expensive.

Using a cheaper model for everything can reduce quality.

A routing system can potentially choose the appropriate model based on the request.

For example:

Simple task → fast low-cost model

Complex reasoning → stronger model

Specialized coding → coding model

Real-time voice → low-latency voice model

Mercury Router is designed around this concept.

If the routing system itself is extremely fast, it can make these decisions without adding significant latency.

Pricing for Mercury 2.5

Inception lists Mercury 2.5 at:

  • $0.20 per million input tokens
  • $0.75 per million output tokens

At launch, the company is offering an 80% discount, bringing the price to:

  • $0.04 per million input tokens
  • $0.15 per million output tokens

The launch pricing is particularly aggressive for developers experimenting with the model.

However, developers should distinguish between the promotional price and the standard listed price.

For production planning, the regular rate is the more important number.

Where Can Developers Use Mercury 2.5?

Inception says Mercury models are available through:

  • Inception API
  • Baseten
  • OpenRouter

Enterprise deployments can also receive dedicated capacity, autoscaling, compliance controls and configurable data-retention options.

Inception is also offering 100 million free tokens for developers trying the API.

That lowers the barrier for developers who want to test whether diffusion-based inference fits their workloads.

How Mercury 2.5 Differs From Large Frontier Models

Mercury 2.5 is not primarily competing by trying to become the largest general-purpose model.

Its strategy is different.

The focus is:

Speed + cost + useful intelligence

That makes it particularly suitable as a component inside larger AI systems.

A developer might use a highly capable frontier model for difficult reasoning while using Mercury 2.5 for:

  • Query rewriting
  • Tool search
  • Context compression
  • Classification
  • Routing
  • Summarization
  • Fast interactive responses

This hybrid approach could become increasingly common.

Why Speed Matters More in Agentic AI

The importance of Mercury 2.5 becomes clearer when looking at AI agents.

A chatbot might generate one response.

An AI agent can perform dozens of model calls.

For example:

User request

↓

Planner

↓

Search

↓

Tool selection

↓

Database lookup

↓

Code execution

↓

Result analysis

↓

Verification

↓

Final response

If every step takes two seconds, the total experience can become frustrating.

A model capable of generating useful responses extremely quickly can reduce the latency of the entire chain.

This is where diffusion LLMs could have a significant advantage.

Diffusion LLMs Could Become More Important

Mercury 2.5 represents a broader shift in the AI model market.

The industry has spent years improving transformer-based autoregressive models.

Now researchers and companies are exploring alternative generation architectures.

Diffusion-based language models are one of those alternatives.

The key attraction is the possibility of generating or refining multiple tokens in parallel.

If the approach continues to improve in reasoning and reliability while maintaining its speed advantage, diffusion models could become an important part of AI infrastructure.

The Biggest Question Is Quality

Speed alone is not enough.

A model generating 1,107 tokens per second is only useful if the output is accurate enough for the application.

That is why Mercury 2.5's claimed 40% intelligence improvement matters.

The company says the model was trained using production failure cases and customer feedback gathered from Mercury 2 deployments.

However, independent testing will be important before developers can determine how Mercury 2.5 compares with competing models across different workloads.

The strongest test will come from real applications.

Mercury 2.5 Could Be Especially Useful as a Supporting Model

One of the most interesting ways to use Mercury 2.5 may not be as the only model in an application.

Instead, it could act as a fast supporting model.

For example:

Frontier model: Handles complex reasoning.

Mercury 2.5: Handles query rewriting, tool selection and context management.

Specialized model: Handles vision, speech or another domain.

This architecture could provide a better balance between quality and cost.

The AI application becomes a team of models rather than a single model.

What Mercury 2.5 Means for Developers

For developers, the launch creates another option when designing AI systems.

Instead of choosing a model solely based on benchmark scores, teams can evaluate:

  • Time to first token
  • Tokens per second
  • Cost per request
  • Context length
  • Tool-calling support
  • Structured output
  • Reasoning controls
  • Hosting options
  • Enterprise controls

Mercury 2.5 is clearly optimized around that broader production equation.

What Mercury 2.5 Means for AI Agents

AI agents could be one of the biggest beneficiaries.

Agents need to perform many small operations quickly.

A model that is extremely fast can make those operations less noticeable to users.

This could make agents feel more responsive.

Search agents could use Mercury for query planning.

Coding agents could use it for context management.

Voice agents could use it for rapid responses.

Enterprise agents could use it for routing and structured tool calls.

The same underlying model can therefore support several types of agentic applications.

Inception Mercury 2.5 is an important AI model launch because it is competing on an axis that is becoming increasingly valuable: how much useful AI work can be completed per second and per dollar.

Inception reports a 40% intelligence improvement over Mercury 2, 1,107 tokens per second on NVIDIA GPUs, a 260K context window, tunable reasoning, parallel tool calls and structured JSON output.

The model is already positioned for search, voice and coding workloads, where repeated model calls can make latency and cost accumulate quickly.

Its diffusion-based architecture makes the launch particularly interesting.

Rather than simply making another autoregressive model, Inception is betting that parallel generation can provide a better foundation for real-time AI applications.

The company is also extending the ecosystem with Mercury Voice and Mercury Router.

That suggests Inception's larger goal is not simply to build one fast LLM.

It is building an infrastructure layer for fast AI applications.

The biggest question now is whether Mercury 2.5's reported intelligence gains hold up across independent evaluations.

If they do, diffusion LLMs could become much more than an experimental alternative.

They could become a serious option for the next generation of search agents, coding assistants, voice systems and real-time AI applications.

FAQs

What is Inception Mercury 2.5?

Mercury 2.5 is Inception's newest production diffusion large language model, designed for low-latency AI applications including search, voice and coding agents.

How fast is Mercury 2.5?

Inception reports a speed of 1,107 tokens per second on widely available NVIDIA GPUs. This is a company-reported performance figure.