Key takeaways

• Mercury 2 is a novel "reasoning diffusion LLM" (dLLM) designed for low-latency coding workflows, real-time voice/search, and agent loops, serving as a fast reasoning LLM.
• The model utilizes a Transformer architecture and generates/refines multiple tokens in parallel, achieving over 1,000 tokens/sec, making it significantly faster than other speed-optimized LLMs.
• Mercury 2 boasts tunable reasoning levels, 128K context length, native tool use, and schema-aligned JSON output, though it is proprietary with closed weights.
• The model excels in academic understanding with 85.97% accuracy on MMLU, particularly in Abstract Algebra (90.0%) and Astronomy (88.1%), but struggles with College Chemistry (70.0%) and nuanced ethical scenarios. It shows strong foundational understanding in mathematical problem-solving with 93.16% accuracy on MATH-500, but struggles with advanced multi-step logical deductions and strict formatting.
• Mercury 2 has a low 26.25% accuracy on Terminal-Bench, performing well on simple coding commands but failing in complex, multi-step instructions, interactive environments, and advanced reasoning tasks. Common failure modes include producing no output for complex tasks, inability to understand interactive environments, and difficulties with tool integration for system configurations.

Technical specifications

Context length

128k tokens

Max output

—

Architecture

Transformer

Modalities

text->text

Parameters

—

Open weights

No

Licence

Proprietary

Regions

USA

Vendor

Inception

Provider model key

inception/mercury-2