Inception MercuryTranslation site

Inception Mercury is a family of language model products designed for developers and enterprises, focusing on rapid text generation, reasoning, and tool invocation. Developed by Inception, an AI research and product company, it offers a web-based chat interface and an API platform. It is ideal for teams building customer service systems, search assistants, code tools, and multi-step agents, and is particularly well-suited for applications with longer generation wait times.Company Profile

Visit the official website - API Platform - Mercury Chat - Development Documentation

What are the features of Mercury?

Mercury adopts a diffuse language model approach, generating answers through multi-round, parallel refinement of text segments. For developers, the key benefit is a reduction in model processing time while retaining familiar interfaces for chat, function calls, and structured data; actual performance still depends on the task, the intensity of reasoning, and the service load.

The Mercury 3, released on October 19, 2026, is the current flagship model. The official generation speed is 1,068 tokens per second, and it runs on an NVIDIA GPU. This figure is based on performance data published by the manufacturer and cannot be directly converted into the first-word time or full-round response time for a specific application.Mercury 3 Release Notes

What models are available on the platform?

• Mercury 3: A general-purpose text reasoning model, identified by the API token "mercury-3." It supports a context size of approximately 1 million tokens, up to 131,072 output tokens, as well as tool calls and structured output. It is suitable for dialogue, information organization, code Q&A, and agent tasks.
• Mercury Voice: A low-latency language model designed for voice-enabled agents that supports text input and output and must be used in conjunction with speech recognition and speech synthesis components. The official documentation states that it is available to enterprise customers; please contact Inception to activate the service.
• Mercury Decide: Specializes in classification, scoring, and decision-making; its API endpoint is /v1/decisions. It accepts text, JSON, and images, and returns structured decisions with probability values.
• Mercury Edit 2: Used for code autocompletion and suggestions for the next editing location, utilizing the FIM and Edit interfaces, respectively—a different division of labor compared to the chat interface in Mercury 3.List of Models and Capabilities

When making a selection, verify the capabilities based on the specific model. The Mercury 3 chat interface only accepts text; image recognition is a capability of Decide and cannot be applied to Mercury 3. The platform also does not provide a public vector embedding or self-service fine-tuning interface. When conducting knowledge base Q&A, vectorization and retrieval require separate configuration.common problems

Key Features and Use Cases

Customer Service, Q&A, and Multi-Step Agents

Mercury 3 can return function calls and JSON parameters; after the application executes a query, it passes the results back to the model to continue the response. It supports parallel tool calls, making it suitable for querying several independent data sources simultaneously. The developer remains responsible for tool permissions, parameter validation, and the actual execution logic.Tool Call Documentation

For example, a customer service representative first looks up an order and then prepares a response; a search assistant reads multiple documents and generates a summary; and a code assistant reads project files and explains errors. When refunds, email communications, or data modifications are involved, clear authorization and confirmation steps should be maintained at the business level.

Structured Data Extraction and Workflow Integration

By constraining output fields using JSON Schema, you can extract work order tags, organize document properties, or generate data for consumption by other programs. Once the format is validated, you should also verify that the field values are accurate, that units are consistent, and that missing values are represented correctly.Structured Output Documents

Real-Time Text and Voice Interaction

Normal streaming mode outputs text in blocks; diffusing mode displays the intermediate text as it is continuously updated. For conversations with real users, it is generally best to use normal streaming output first to avoid having text that is still changing read aloud. Voice applications can use the “speech recognition → Mercury → speech synthesis” pipeline.Streaming Output Instructions,Voice Application Guide

How do I get started?

1. Go to the Inception Platform to create an account. According to the official Quick Start guide, new accounts currently receive a one-time allocation of 100 million free tokens, and you don’t need to provide payment information to start the trial. This quota is shared among all models in the account; it is not reissued each time a new secret key is created.
2. Create an API key in your account and store it in your project's environment variables or a key management system. Do not include it in your web front end, public repositories, or shared screenshots.
3. Use the official Python or TypeScript SDK, or a client that supports the OpenAI-compatible interface. The base API URL is https://api.inceptionlabs.ai/v1; send chat requests to /chat/completions and specify "mercury-3" as the model.
4. Start by testing `messages`, `output`, and `usage` with short tasks that do not contain sensitive information. Enable `stream` when you need to display results as they are generated; set `max_completion_tokens` according to the task to avoid running out of output budget.
5. Explicitly select `reasoning_effort`, then evaluate speed and quality. Mercury 3 offers `instant`, `low`, `medium`, and `high`; if not specified, the default is `low`. Start by comparing results using a single fixed configuration to avoid mixing different reasoning intensities within the same set of results.
6. Before going live, check account limits, error handling, and logs. If a 429 error occurs, throttle the request, back off, and retry; set timeouts, retry limits, and manual fallback paths for service overloads and business tool failures.Quick Start,Inference Strength,Error Handling

Fees and Limits

The following are the unit prices for the Mercury 3 direct API, as retrieved on October 20, 2026, in U.S. dollars per million tokens.

The official website states a 50% discount for new users, but the press release and pricing page I’ve reviewed do not specify an end date for the promotion. When budgeting, you should also consider the regular price and double-check your account statement before making a purchase. Once your free allowance is used up, you’ll need to set up a payment method; enterprise-level capacity, limits, and service guarantees are subject to separate negotiation.Official Pricing Page

Cache hits are billed at the cache rate; reasoning tokens are also included in the output cost. When estimating your bill, do not add the `reasoning_tokens` listed under "Usage" again, as they are already included in the `completion_tokens`. Pricing and quotas on third-party platforms may differ from those of direct-connect accounts.Billing Criteria

How does it compare to other tools?

The following table compares development platforms and specific available models to help you select the right one for your task. All prices are standard online unit prices publicly available at the time of inquiry; since different platforms have varying billing structures for tokenization, inference volume, and additional tools, this table is not a substitute for conducting a cost analysis for a specific task.

Basis for Comparison:Groq Model List,Groq Compatibility Notes,Gemini 3.5 Flash-Lite Model Description,Gemini API Pricing.

If the input consists primarily of text and your application already has a dedicated toolset, you can start by comparing the task success rates and end-to-end latency of Mercury and Groq. If you need to process images, audio, video, or PDFs directly, you should first evaluate the input capabilities of multimodal models such as Gemini before assessing their cost and performance.

What should I keep in mind before connecting?

• Platform Requirements: A web browser is required to access the web interface; using the API requires an account, a secret key, and a network connection capable of accessing the service endpoints. The official documentation provides integration guides for Python, TypeScript, and common frameworks.
• Regional Availability: The page reviewed does not provide a sufficiently comprehensive list of regions where Inception is supported; therefore, we cannot guarantee that registration, payment, or API calls will be available in Mainland China based on this information. This should be confirmed in the actual deployment region; Gemini has separate regional and account requirements.
• Data Handling: Do not assume that standard accounts have zero data retention by default. For sensitive operations, you should first confirm retention, access, and corporate agreement requirements with the service provider.Inception Data Processing Guidelines,Gemini Regional Requirements
• Migration Arrangements: Mercury 2 and 2.5 are no longer available to new users; customers with a usage history prior to October 19, 2026, will continue to have access until November 15, while enterprise customers will follow the terms of their agreements. "mercury-2.5" may still appear in legacy examples; new projects should follow the Mercury 3 Quick Start guide. During migration, re-confirm the inference strength and perform regression testing.Official Migration Notice

When selecting a model, you should test the accuracy of Chinese text processing, recall for long documents, the accuracy of tool invocations, and the P95 latency during peak business hours.