Breaking

Inception Adds Diffusion-Based LLM Mercury 2 to Design Arena

June 29, 2026 at 18:19 EDT

Foundation Models
AI Agents
Infra & Chips

Mercury 2, a diffusion-based large language model developed by Inception Labs (@_inception_ai), is now available on Design Arena, a design-oriented benchmark platform. The model is pitched on delivering reasoning-grade quality while generating fast through parallel refinement.

February 2026 · Inception Labs

Mercury 2: The World's First Reasoning Diffusion LLM

A diffusion-based model that refines many tokens in parallel — chasing reasoning-grade quality at roughly 1,000 tokens/sec, now live on Design Arena.

1,009 tok/s

Official speed on NVIDIA Blackwell

128K

Context length

$0.75/1M

Output price ($0.25 input)

Partner-reported production gains

Reduction figures from summarization workloads — taller stack = bigger cut.

82%

Latency reduction

90%

Cost reduction

How it generates text differently

Conventional

One token at a time, left to right

Mercury 2 (dLLM)

Many tokens refined in parallel from noise

Quality benchmarked comparable to Claude 3.5/4.5 Haiku and GPT-5.2 Mini · OpenAI API compatible.

Strengths

Near-1,000 tok/s — wait time all but vanishes
Smooth agent loops, voice, and code editing
Low cost; native tool use & JSON output

Drawbacks

Writing quality seen as medium to poor
Hallucinations, fabricated references, slop
Weak emotional nuance; feels slower than gen 1

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.