😸 Diffusion models are coming for text at ~$0.80 per MILLION flat

2026-02-24 · Source: The Neuron · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Intermediate, long

Summary

Inception Labs has launched Mercury 2, a new diffusion LLM designed for text editing rather than generation, operating at 1,196 tokens per second, which is over three times faster than its closest competitors like Claude 4.5 Haiku (~89 tokens/sec) and GPT-5 Mini (~73 tokens/sec). This model starts with a full draft and refines it in parallel, similar to how AI image generators function. Mercury 2 is priced at $0.25 per million input tokens and $0.75 per million output tokens, making its output cheaper than GPT-5 Mini. It ranks 18th out of 134 models on Artificial Analysis's intelligence index, excelling in agentic coding and instruction-following, with support for tool use, 128K context, and OpenAI-compatible structured outputs. The article also highlights other AI developments, including DeepSeek's alleged use of banned NVIDIA Blackwell chips and distilled knowledge from major AI models, Anthropic's new enterprise plugins for Cowork, and significant investments in AI chip startups like MatX ($500M) and a $100B deal between Meta and AMD for Instinct compute.

Key takeaway

For Machine Learning Engineers building agent-based systems, Mercury 2's 10x speed advantage fundamentally changes what is possible by drastically reducing latency in chained AI calls. You should evaluate integrating diffusion LLMs like Mercury 2 into your workflows to enable more natural voice assistants, faster code agents, and background automations that complete tasks efficiently, especially given its competitive pricing and OpenAI-compatible stack.

Key insights

Diffusion models can significantly accelerate text processing by editing full drafts in parallel, rather than generating word-by-word.

Principles

• Parallel refinement enhances AI text processing speed.
• Diffusion models are adaptable beyond image generation.
• Agent loops benefit disproportionately from reduced AI latency.

Method

Mercury 2 employs a diffusion LLM approach, starting with a rough sketch of an entire answer and then refining it all at once, mimicking an editor's parallel revision process for text and reasoning tasks.

In practice

• Utilize Mercury 2 for high-speed agentic coding.
• Implement Anthropic's Cowork plugins for role-specific AI tasks.
• Explore Qwen3.5 for local, frontier-class model deployment on Mac.

Topics

• Diffusion LLMs
• AI Agent Workflows
• AI Hardware
• Enterprise AI
• AI Geopolitics

Code references

• qwibitai/nanoclaw
• anthropics/knowledge-work-plugins
• generalaction/emdash

Best for: CTO, Machine Learning Engineer, NLP Engineer, AI Engineer, AI Product Manager, Investor

Related on AIssential

• Towards Speed-of-Light Text Generation with Nemotron-Labs Diffusion Language Models — Diffusion Language Models (DLMs) enable parallel, iterative text generation and revision, significantly boosting speed and accuracy over autoregressive models.
• GLM-5.2, NEW Gemini Model, Mira Murati's New Model, Fable 5 Update, & More! AI NEWS — The AI landscape is rapidly evolving with new models focusing on context, speed, and real-time interaction, alongside strategic competitive moves.
• True Positive Weekly #165 — This brief highlights key advancements in AI, data science, and practical GPU optimization techniques.
• 😺 🎙️ Mercury 2: AI that's 10x faster than ChatGPT & Claude — Diffusion models offer a 10x speedup and cost reduction over autoregressive LLMs by processing tokens in parallel.
• The Race to Production-Grade Diffusion LLMs with Stefano Ermon - #764 — Diffusion language models offer significantly faster and more efficient inference for text and code generation than autoregressive LLMs.
• Mercury 2: The AI Model That Feels Instant — Diffusion-based language models offer superior speed and cost efficiency over auto-regressive models for complex reasoning tasks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Neuron.