Mercury 3

Inception

Mercury 3 is Inception's flagship diffusion-based LLM (dLLM), released 2026-10-19 and described as the most capable diffusion LLM released to date. Inception says it is more than 40x larger in total parameters than Mercury 1 and runs at about 1,068 output tokens per second on NVIDIA GPUs while matching or beating speed-optimized models such as GPT-6 Luna and Claude Haiku 5.5. It has a 1M-token context window, up to 131,072 output tokens, tunable reasoning effort (instant, low, medium, high), tool calling, and structured outputs. Text in, text out. Available on the Inception API, OpenRouter, Vercel AI Gateway, and Baseten.

Key Specifications

Parameters
-
Context
1.0M
Release Date
October 19, 2026
Average Score
60.1%

Timeline

Key dates in the model's history
Announcement
October 19, 2026
Last Update
October 27, 2026
Today
November 20, 2026

Technical Specifications

Parameters
-
Training Tokens
-
Knowledge Cutoff
-
Family
-
Capabilities
MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)
$0.10
Output (per 1M tokens)
$0.35
Max Input Tokens
1.0M
Max Output Tokens
131.1K
Supported Features
Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning

Benchmark Results

Model performance metrics across various tests and benchmarks

Reasoning

Logical reasoning and analysis
GPQA
Vendor benchmark table (Mercury 3, medium reasoning effort); GPQA Diamond. • Self-reported
89.2%

Other Tests

Specialized benchmarks
AA-LCR
Vendor benchmark table (Mercury 3, medium reasoning effort); AA-LCR v1.1. • Self-reported
77.7%
BrowseComp-Plus
Vendor benchmark table (Mercury 3, medium reasoning effort); BrowseComp+. • Self-reported
55.2%
DeepSearchQA
Vendor benchmark table (Mercury 3, medium reasoning effort); DeepSearchQA. • Self-reported
71.3%
DeepSWE 1.0
Vendor benchmark table (Mercury 3, medium reasoning effort); DeepSWE v1. • Self-reported
46.3%
GDPval-AA
Vendor benchmark table (Mercury 3, medium reasoning effort); GDPval-AA v2 Elo (max_score 3000). • Self-reported
41.2%
IFBench
Vendor benchmark table (Mercury 3, medium reasoning effort); IFBench. • Self-reported
76.8%
MRCR v2
Vendor benchmark table (Mercury 3, medium reasoning effort); MRCR v2; needle count and context length not specified. • Self-reported
75.5%
Multi-Challenge
Vendor benchmark table (Mercury 3, medium reasoning effort); MultiChallenge. • Self-reported
60.4%
OmniScience
Vendor benchmark table (Mercury 3, medium reasoning effort); AA-Omniscience (accuracy). • Self-reported
46.4%
SWE-Atlas
Vendor benchmark table (Mercury 3, medium reasoning effort); Labeled 'SWE-Bench Atlas' in vendor table; split and harness not specified. • Self-reported
51.1%
Tau3 Banking
Vendor benchmark table (Mercury 3, medium reasoning effort); τ³-bench Banking. • Self-reported
30.2%

License & Metadata

License
proprietary
Announcement Date
October 19, 2026
Last Updated
October 27, 2026

Compare Mercury 3

All comparisons

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.