Inception Platform (Mercury LLM)

Inception Platform provides production access to Mercury LLM, Inception's family of diffusion large language models, through managed APIs and enterprise deployment arrangements. It serves developers and enterprise teams building conversational applications, coding tools, and workflows that generate text, call functions, or return structured decisions. Its core offering is AI Model Serving & Inference, with configurable data handling, capacity, and support terms.

Mercury LLM is the core of the platform. Where most language models write one token at a time, Mercury is a diffusion model: it starts from a rough draft of the whole response and refines many tokens in parallel. Mercury 3, the latest flagship model, supports a 1 million token context window and up to 131,072 output tokens. It offers tool calling, JSON Schema structured outputs, and four reasoning effort levels (instant, low, medium, high), and accepts text input. Inception reports speeds above 1,000 tokens per second on NVIDIA GPUs, a figure the company measured itself. The family also includes Mercury Edit 2 for code completion and editing, Mercury Decide for classification and extraction, and Mercury Voice for voice agents, which is available to enterprise customers through sales. Earlier models, Mercury 2.5 and Mercury 2, remain available.

Key Capabilities

• Diffusion-based generation: Mercury iteratively refines text during generation. With stream: true, the API returns refined blocks of output instead of single tokens. Adding diffusing: true exposes the intermediate refinement steps, and each chunk then contains the full text at that step.
• Reasoning and tool use: Developers set reasoning effort per chat request, from instant for voice and high-volume automation to high for multi-step tasks. Mercury 3 defaults to low. Developers define functions the model can call, the surrounding application executes them, and the results return to the model to continue the workflow.
• Structured responses and decisions: Chat models support JSON Schema output. Mercury Decide has a separate endpoint for classification, categorical choices, ordinal scores, and extraction of exact text spans with offsets and confidence, over shared context. It accepts text, JSON, and image input.
• Code completion and editing: Mercury Edit 2 supports fill-in-the-middle completion using code before and after the cursor, plus next-edit predictions informed by file context, recently viewed snippets, and edit history.
• Enterprise controls: Inception offers configurable retention and caching, with private networking, dedicated capacity, throughput commitments, and support terms available according to the deployment agreement. It states that customer prompts and outputs are not used for training.
• Integrations: Documented integrations and cookbooks cover Claude Code, Cursor, Zed, Kilo Code, OpenCode, LangChain, ElevenLabs, and Vapi.

Audience & Use Cases

Inception Enterprise serves software vendors and internal development teams embedding models in search, support, coding, and business applications. Voice-agent developers can use Mercury as the language-model component of a speech application, while editor developers can integrate its completion and editing endpoints. The documentation includes worked examples for search agents, natural-language SQL agents, coding agents, and voice customer support.

Technical Specifications

The service exposes an OpenAI-compatible API and official Python and TypeScript clients. Model-specific endpoints cover chat, code completion, code editing, and decisions. Mercury 3 supports a 1M-token context window and up to 131,072 output tokens. Mercury 2.5 supports 260K tokens of context and up to 65,536 output tokens. Mercury Edit 2 supports 32K tokens of context and up to 8,192 output tokens. Access is hosted, with Free and Pay As You Go tiers that have per-minute request and token limits. Enterprise capacity, networking, and commercial requirements are negotiated.