Inception Labs' diffusion language model — parallel token refinement at 1,000+ tokens per second with a 1M-token context, built for latency-sensitive agents.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "mercury-3",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Overview
Mercury 3 is Inception Labs' flagship diffusion-based language model, released in October 2026. Instead of generating tokens one at a time, it produces and refines them in parallel, reaching 1,000+ tokens per second with a 1M-token context window — built for agents, search, and real-time applications where latency compounds.
On Venice, Mercury 3 runs under the anonymized privacy tier: your prompts are not stored, profiled, or used for training, and there is no personal generation history tied to your account. Access is permissionless — pay per token through the Venice app or the OpenAI-compatible API, no subscription required. That makes the fastest closed model in its price class usable without feeding a Big-Tech data pipeline.
Agent quickstart
The API is OpenAI-compatible: change the base URL and the model id and existing client code works unchanged.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "mercury-3",
"stream": true,
"messages": [{ "role": "user", "content": "Draft the release note." }]
}'curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "mercury-3",
"messages": [{ "role": "user", "content": "Find the rate limits." }],
"tools": [{
"type": "function",
"function": {
"name": "search_docs",
"description": "Search the API documentation.",
"parameters": {
"type": "object",
"properties": { "query": { "type": "string" } },
"required": ["query"]
}
}
}],
"tool_choice": "auto"
}'import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["VENICE_API_KEY"],
base_url="https://api.venice.ai/api/v1",
)
resp = client.chat.completions.create(
model="mercury-3",
messages=[{"role": "user", "content": "Build without permission."}],
)
print(resp.choices[0].message.content)Specifications
Assessment
Use cases
Prompting
Enable function calling and pass your tool schemas up front — Mercury 3 is built for agentic loops where round-trips dominate latency.
Use structured output with a JSON schema to get parseable results in one shot instead of re-rolling freeform text.
Turn on web search for questions about current events rather than relying on parametric knowledge.
Put stable instructions in a cached prefix — cached input bills at $0.01 per 1M tokens, so repeated system prompts cost almost nothing.
For coding subagents, keep the reasoning setting on for hard steps and off for mechanical ones to trade quality against the already-fast decode.
Stream the output in your UI — at 1,000+ tokens per second, non-streamed responses arrive faster than users can read them anyway.
Alternatives
| Model | Best for | Context | Open weights | Price (Venice) |
|---|---|---|---|---|
| Mercury 3 | Latency-sensitive agents & real-time apps | 1M tokens | No | $0.10 in · $0.35 out / 1M |
| Kimi K3 | Open-weight frontier reasoning | 1M tokens | Yes | $3.75 in · $18.75 out / 1M |
| Gemini 3.8 Flash | Fast general-purpose chat | 1M tokens | No | $0.94 in · $4.69 out / 1M |
| Qwen 3.7 Plus | Balanced long-context work | 1M tokens | No | $0.50 in · $2 out / 1M |
Pick Mercury 3 when wall-clock time is the product — agents, voice, live search — and you want speed-tier quality at the lowest output price on the board. Pick Kimi K3 instead when owning the weights matters more than milliseconds.
Pricing
Billed per token on Venice: $0.10 per 1M input tokens and $0.35 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Getting a key
Create a key in API settings. Nothing else is required to start.
Export it as VENICE_API_KEY so the snippets above run unedited.
Point an existing OpenAI client at https://api.venice.ai/api/v1. The scheme is part of the value: an OpenAI client given a bare host does not resolve it.
Pass mercury-3 as the model and send the request.
FAQ
Mercury 3 is Inception Labs' flagship diffusion-based language model, released in October 2026. It generates text by producing and refining tokens in parallel rather than one at a time, reaching over 1,000 tokens per second with a 1M-token context window and a 128K-token output ceiling.
Venice bills Mercury 3 per token: $0.10 per 1M input tokens, $0.35 per 1M output tokens, and just $0.01 per 1M for cached input. Note that Inception's launch pricing is a 50% discount off a reported standard rate of $0.20 in / $0.70 out per 1M tokens.
Neither. Mercury 3 is a closed, proprietary model — Inception has not published the weights, so it cannot be self-hosted or fine-tuned. On Venice you pay per token with no subscription required, and new accounts include free daily prompts and welcome credits to try it.
They trade on ownership versus speed. Kimi K3 publishes open weights under a Modified MIT licence, so you can download, inspect, and self-host it — but it costs $3.75 in / $18.75 out per 1M tokens. Mercury 3 is closed and rental-only, yet generates at 1,000+ tokens per second for $0.10 in / $0.35 out. Choose Kimi for sovereignty, Mercury 3 for latency and cost.
Inception Labs measures Mercury 3 at 1,068 tokens per second on NVIDIA GPUs, and OpenRouter confirmed 1,000+ tok/s at launch — roughly 4x the throughput of Mercury 2.5. These are vendor and platform figures rather than independent benchmarks, but the diffusion architecture's parallel refinement is the reason for the speed.
Yes. On Venice, Mercury 3 supports function calling, a reasoning mode, web search grounding, and structured output via JSON schema — the full toolkit for agentic workflows. It does not, however, accept image or audio input; it is a text-only model.
Venice runs Mercury 3 under its anonymized privacy tier: prompts are not stored, profiled, or used for training, and requests carry no personal account history. Unlike some Venice models, it does not run inside a trusted execution environment with end-to-end encryption, but zero retention of prompts still applies.
Mercury 3 has a 1M-token context window with a 128K-token maximum output, a major expansion from the roughly 260K-token context of its predecessor. That makes it practical for whole-codebase reasoning and very long agent trajectories.
Mercury 3 was released by Inception Labs in October 2026 and went live on Venice the same month. Some listings date the launch to October 16 and others to October 19, 2026; either way, it is available to use on Venice today.
One key, free to start, no credit card.