Introduction
Mercury 3 from Inception supports responsive text reasoning and agent interactions. This chat API accepts text, provides a 1,000,000-token input budget and accepts max_tokens up to 131,072. Stream Mercury 3’s text as it arrives, or let a function-call result supply the evidence for the next step in fast text turns with Mercury 3.
| Field | Value |
|---|---|
| Input | Text |
| Output | Text |
| Input budget | 1,000,000 tokens |
| max_tokens ceiling | 131,072 |
| Inputs | text |
Key features & highlights
Written in parallel, not one token at a time
Most language models write left to right, one token after another. Mercury is a diffusion model: according to Inception it produces and refines many tokens at once, the way image diffusion models refine a whole picture. Inception presents this as the reason Mercury 3 can run at over 1,000 tokens per second.
Built for the places where waiting hurts
Inception reports 1,068 tokens per second on NVIDIA GPUs and describes Mercury 3 as 4-8x faster than comparable models. These are the publisher's own measurements, so treat them as a direction rather than a promise: speed on any API depends on load, prompt length and reasoning effort. Voice, support chat, search agents and browser agents are the cases Inception names.
A million tokens of input
1,000,000 tokens is roughly 750,000 English words, up from 260K in Mercury 2.5 according to Inception. A long thread, a manual or a repository can travel in the request as written, so the model reads the original rather than a summary.
Reasoning you can turn down for speed
reasoning_effort takes none, low, medium or high, and low is the default here. Keep it at none or low for tagging, routing and quick replies, and raise it when a task needs several steps. Tools, streaming and JSON output work on the same call.
Request parameters
| Parameter | Required | Description |
|---|---|---|
messages | Yes | For fast text turns with Mercury 3, supply the conversation and the evidence as text in order, and connect each tool result to its original |
max_tokens | No | Budget Mercury 3’s visible answer and billable reasoning together in |
reasoning_effort | No | Mercury 3 exposes none, low, medium, high, starting at low; compare effort settings on fast text turns with Mercury 3 because deeper reasoning can increase output tokens and wai… |
stop | No | End Mercury 3’s generated text at one supplied stop sequence or any of up to four; the matching sequence itself is left out of the answer. at least 1 characters · at least 1 items · at most 4 items |
temperature | No | Mercury 3 accepts temperature from 0.5 to 1; keep other sampling controls steady when comparing wording for fast text turns with Mercury 3. |
top_p | No | Use |
tools | No | Give Mercury 3 OpenAI-format function definitions for fast text turns with Mercury 3; the model proposes arguments and your application validates and executes the operation. at most 128 items |
parallel_tool_calls | No | Allow Mercury 3 to propose independent function calls together when fast text turns with Mercury 3 needs more than one external result. Default: |
tool_choice | No | Choose auto, none, required or a specific function for Mercury 3, then preserve the returned call ID when continuing fast text turns with Mercury 3. |
response_format | No | Ask Mercury 3 for text or structured JSON, then validate the JSON fields needed for fast text turns with Mercury 3 before your application uses them. |
Pricing
Input $0.20 / Output $0.70 per 1 million tokens
Use cases
Real-time support and voice assistants
When a customer is waiting, time to the first useful sentence matters. Stream the reply, keep reasoning_effort at none or low, and put the tone and policy in the system message.
Search and browser agents
An agent that calls a tool, reads the result and calls another can spend most of its time waiting on the model. Give Mercury 3 your search or page-reading functions and let it run the loop. Inception names browser-use and real-time search agents among its early partners.
High-volume classification
Label messages, route tickets or pull fields out of text at scale. Ask for a JSON object, set a low temperature and parse the result directly, with no post-processing.
Drafting and rewriting on the fly
Short replies, summaries, tutoring feedback and rewrites feel closer to instant when generation is fast. For long outputs, raise max_tokens and stream the text as it arrives.
Prompt guide & pro tips
Keep reasoning low for latency-sensitive paths
The default is low. Use none for the quickest turns, and move to medium or high only for tasks that need planning. Compare a handful of your own prompts at each level before you pick a default.
Stay inside the temperature range
This endpoint accepts temperature from 0.5 to 1. A lower value suits classification and extraction; a higher one gives more varied wording in drafts.
Set
max_tokensfor the answer you wantThe default
max_tokensis 4,096 and the ceiling is 131,072. Thinking, when used, is counted against the same budget, so leave headroom at higher effort levels and check the finish reason.Ask for structure when code will read the answer
Use
response_formatto request a JSON object, or define a tool whose parameters are the shape you need, and validate what comes back before acting on it.
Low-latency text work
Mercury 3 uses diffusion to refine text in parallel. For a support reply, send the customer’s question, the relevant policy excerpt and the desired next action as text; this endpoint does not accept image, video or audio parts.
Keep the interaction responsive
Start at low reasoning effort, then compare none for simple labeling tasks. Stream long replies and reserve enough max_tokens for reasoning as well as the visible answer. Inception’s published speed is a measurement of its own setup, not a latency promise for your request.
Connect an external action
Mercury 3 can propose a function call; your application checks the arguments and performs it. Return the result with the matching tool_call_id so the next quick turn can use evidence from the operation. Review the final answer against the supplied policy before showing it to a customer.
Mercury 3’s own content decisions still apply during fast text turns with Mercury 3; NSFW responses may change with the request or model version, so unrestricted output is not promised.
API quick start
curl --fail-with-body -sN https://api.spicyapi.ai/v1/chat/completions \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $(uuidgen)" \
-d '{
"model": "inception/mercury-3/chat",
"messages": [
{
"role": "user",
"content": "Summarise the following release notes into three bullet points."
}
],
"stream": true
}' \
| jq --unbuffered -Rrj 'select(startswith("data: ")) | .[6:] | select(. != "[DONE]") | fromjson | if .error then "\n[error] \(.error.message)\n" else (.choices[0].delta.content? // empty) end'
See the API reference for authentication, the complete schema, polling, and error handling.