Mercury 2
Mercury 2 is the world's fastest reasoning LLM, built on diffusion architecture for 5x faster generation and real-time agent workflows.
Get API keyWhat is Mercury 2?
Mercury 2 is a diffusion-based reasoning language model developed by Inception Labs, released on February 24, 2026. It generates text through parallel refinement instead of autoregressive decoding, achieving over 1,000 tokens per second and enabling ultra-low latency for agent loops, real-time search, and coding workflows.
Use Mercury 2 privately on Venice
On Venice, Mercury 2 runs with anonymized privacy — your prompts are never stored or profiled. This means you get the speed and reasoning capabilities of a cutting-edge dLLM without sacrificing sovereignty. The model is uncensored in operation on Venice, supporting permissionless deployment of agent pipelines and structured automation.
What can Mercury 2 do?
- •World's fastest reasoning LLM — achieves over 1,000 tokens/sec via parallel refinement, making it ideal for latency-sensitive production workflows.
- •5x faster than leading speed-optimized autoregressive models like Claude Haiku 4.5 and GPT-5 Mini, drastically reducing cumulative latency in agent loops.
- •Supports tunable reasoning depth, tool use, and schema-aligned JSON output for structured automation.
- •Cost-efficient at $0.31/$0.94 per million tokens (in/out), with significant savings via prompt caching.
- •128K context window enables long-form reasoning and document processing.
- •Lower reasoning quality compared to frontier models like Claude Opus or GPT-4 — GPQA score of 77.0% places it below top-tier reasoning benchmarks.
- •Not open-source or open-weights, so self-hosting and fine-tuning are not possible.
- •Specialized architecture may not benefit all workloads — gains are most pronounced in high-volume, low-latency agent pipelines.
Mercury 2 capabilities
- Tool use / function calling
- Vision (image input)
- Reasoning
- Web search
- Code-optimized
- Structured output (JSON schema)
- Audio input
- Video input
- Multiple image inputs
- Log probabilities
How to use Mercury 2 via API
Venice exposes an OpenAI-compatible API. Swap your base URL and call mercury-2.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "mercury-2",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Specifications
Pricing
Billed per token on Venice: $0.31 per 1M input tokens and $0.94 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Mercury 2 vs alternatives
| Model | Speed (tok/s) | Input Price /1M | Output Price /1M | Open weights |
|---|---|---|---|---|
| Mercury 2 | 1,009 | $0.31 in · $0.94 out / 1M | $0.94 | No |
| DeepSeek V4 Flash 0731 | 850 | $0.17 in · $0.35 out / 1M | $0.35 | No |
| Claude Sonnet 4.6 | 320 | $3.60 in · $18 out / 1M | $18.00 | No |
| Google Gemma 4 31B Instruct | 410 | $0.12 in · $0.36 out / 1M | $0.36 | Yes |
World's fastest reasoning LLM via diffusion; 5x faster than autoregressive peers.
What is Mercury 2 good for?
- •High-throughput agent loops where latency compounds across multiple LLM calls.
- •Real-time voice and search applications requiring instant responses.
- •Lightweight coding assistants and interactive editing workflows.
- •Form processing and structured data extraction with JSON output.
- •Cost-sensitive production AI deployments needing speed without frontier-level reasoning.
Prompting tips
- •Use concise, directive prompts — Mercury 2 excels at fast, deterministic tasks rather than exploratory reasoning.
- •Leverage tool calling and JSON schema for reliable automation pipelines.
- •Break long tasks into smaller steps to maximize throughput and minimize timeouts.
- •Use caching strategically — repeated context is billed at $0.03/M tokens, cutting costs by up to 80%.
Version history
Predecessor model with autoregressive architecture.
CurrentCurrent — diffusion-based, 5x faster, production-grade reasoning.
Frequently asked questions
Mercury 2 is a diffusion-based reasoning language model developed by Inception Labs, released on February 24, 2026. It uses parallel refinement instead of autoregressive decoding to generate text, achieving over 1,000 tokens per second — making it the fastest reasoning LLM available for production AI workflows.
On Venice, Mercury 2 costs $0.31 per million input tokens and $0.94 per million output tokens. Cached input is billed at just $0.03 per million tokens, making repeated context highly cost-efficient.
No. Mercury 2 is a proprietary model developed by Inception Labs and is not available as open weights or open source. It cannot be self-hosted or fine-tuned.
Yes. Mercury 2 supports native function calling and tool use, making it well-suited for agent workflows, retrieval pipelines, and structured automation tasks.
Yes. Mercury 2 has built-in web search capability, enabling it to perform real-time lookups and multi-hop retrieval tasks as part of its reasoning process.
Mercury 2 is significantly faster (1,009 vs 320 tokens/sec) and much cheaper per token than Claude Sonnet 4.6, making it superior for high-volume, low-latency agent loops. However, Sonnet offers stronger reasoning quality for complex tasks. Choose Mercury 2 for speed and cost, Sonnet for depth.
Mercury 2 has a 128K token context window, allowing it to process long documents, codebases, or multi-step agent histories in a single pass.
Yes. Mercury 2 supports structured output via JSON schema, enabling reliable data extraction and integration with automation pipelines.
Related models
Run Mercury 2 privately.
No prompt logging. No data used for training. Free to start — no credit card.
