# Diffusion LLMs Are Here! Is This the End of Transformers?

Source: https://www.youtube.com/watch?v=Bqdf6Um_8OE
Recap page: https://rapidrecap.app/video/Bqdf6Um_8OE
Generated: 2026-02-24T18:36:09.861+00:00

---
## Quick Overview

Diffusion LLMs, exemplified by Inception's Mercury 2 model, are demonstrably faster than traditional autoregressive LLMs in inference speed, completing tasks like writing code or generating text significantly quicker by processing tokens in parallel blocks rather than sequentially, although autoregressive models like GPT-4 still hold an edge in complex reasoning benchmarks.

**Key Points:**
- Inference speed comparisons show Inception Diffusion LLMs (like Mercury 2) completing tasks (e.g., writing a function) significantly faster than Autoregressive LLMs, taking 14 iterations versus 75 iterations for the same task (00:06:07).
- Diffusion models generate tokens in parallel blocks, which is inherently faster than the serial, left-to-right generation of autoregressive models, resulting in Mercury 2 being up to 10 times faster in certain inference scenarios (00:12:15).
- The diffusion approach allows for iterative refinement of the entire sequence simultaneously, overcoming the 'Errors Cascade' problem inherent in sequential generation where early mistakes compound (01:46:18).
- In a coding benchmark comparing Mercury 2 against leading models, Mercury 2 achieves an Average Quality score of 68.3, outperforming several models like Gemini 2.5 Flash (57) and Claude 3.5 Haiku (62) in this specific metric (11:07).
- Despite the speed advantage, the benchmark chart shows that models optimized for coding, like Mercury Coder Small, occupy the 'Artificial Analysis' quadrant by achieving high coding index scores while maintaining high output speeds (~1008 tokens per second) (11:46).
- Inception offers three Mercury models (Mercury, Coder, and 2) with varying capabilities and pricing; Mercury 2 costs $0.75 per million output tokens, includes reasoning effort, and supports streaming, tool use, and structured output (12:21).
- The diffusion architecture enables complex real-time applications, such as rapidly processing a voice query about store hours in real-time, which the model handles instantly (08:51).

![Screenshot at 00:06: 07:A side-by-side comparison shows the Diffusion LLM completing the task in 14 iterations while the Autoregressive LLM requires 75 iterations, visually demonstrating the significant speed advantage of the diffusion approach.](https://ss.rapidrecap.app/screens/Bqdf6Um_8OE/00-00-06.jpg)

**Context:** The video introduces and benchmarks Diffusion Large Language Models (LLMs) developed by Inception, specifically focusing on the Mercury model family, as an alternative to traditional autoregressive transformer models like GPT and Claude. The core argument centers on the superior inference speed of diffusion-based generation, which processes tokens in parallel blocks, contrasting this with the slow, serial nature of left-to-right autoregressive decoding. The presentation includes demonstrations of code generation, complex story writing with length constraints, and formal quality benchmark comparisons against leading models.

## Detailed Analysis

The video proves that Diffusion LLMs, such as Inception's Mercury models, offer a massive speed improvement over autoregressive LLMs for inference tasks. A demonstration of writing an LLM inference function shows the Diffusion model completing in 14 iterations compared to 75 iterations for the Autoregressive model (00:06:07). This speed stems from the diffusion architecture's ability to generate tokens in parallel blocks (coarse-to-fine generation) rather than sequentially (left-to-right generation), as illustrated by the function writing comparison (00:00:00). This parallel processing also allows the model to correct errors across the entire sequence simultaneously, mitigating the compounding error issue found in sequential models (01:46:18). Furthermore, the models demonstrate strong capability in complex tasks, such as writing a story with strict word count progression up and down (04:51), and successfully executing Retrieval-Augmented Generation (RAG) tasks like finding specific data in documents in 1.58 seconds (10:16). Benchmark data supports these claims, showing Mercury 2 achieving an average quality score of 68.3, surpassing several competitors on the chart, although models like Claude 4.5 Haiku (Reasoning) score higher in quality (62 vs 68.3) but are significantly slower (62.6 seconds latency vs. Mercury 2's undefined latency) (11:07). Inception also highlights its competitive pricing structure for its models (Mercury, Coder, 2) and emphasizes the real-time capability of the diffusion approach through demonstrations involving voice interaction (07:51) and complex document retrieval (09:31).

### LLM Inference Speed Comparison

- Autoregressive LLM (Left-to-Right Generation) vs. Inception Diffusion LLM (Coarse-to-Fine Generation)
- Function writing task takes 75 iterations for Autoregressive vs. 14 iterations for Diffusion (00:06:07)
- Diffusion is shown to be up to 10 times faster in inference (00:13:13)

### The Problem with Autoregressive Models

- Errors Cascade
- Sequential token commitment means early mistakes compound, demonstrated with 'The quick brown cat jumps...' where 'cat' forces incorrect subsequent tokens (01:46:18)
- Diffusion models refine all tokens simultaneously, acting like an editor to fix mistakes across the output (01:57:20)

### Code Generation & RAG Capabilities

- Tetris Game Generation
- GPT-4 (Instant mode) and Mercury 2 both successfully generate a complex HTML/JS Tetris game where blocks fall up, with both implementations appearing functional (03:47:00)
- RAG Example: Answering a question about an acquisition agreement purchase price took 2 steps in 1.58s using Mercury 2 (10:03)

### Quality Benchmark Comparison (Mercury 2 vs. Leading Models)

- Mercury 2 achieves an Avg. Quality of 68.3
- It outperforms Gemini 2.5 Flash (57) and Claude 4.5 Haiku (62) but is outperformed by Claude 4.5 Haiku (Reasoning) (62) and GPT-5 (51) in overall quality (11:07)
- Mercury 2 excels in the 'Artificial Analysis' quadrant on the Coding Index vs. Output Speed chart, showing high quality/speed balance (11:46)

### Mercury Model Family Pricing & Features

- Mercury 2 Output Price is $0.75 per 1M tokens, lower than Mercury ($1.00)
- Mercury 2 includes Reasoning Effort capability, which is marked as unavailable ('/') for Mercury and Mercury Coder (12:31)

![Screenshot at 00:01: 01:Code comparison showing the Autoregressive LLM function requiring sequential token generation versus the Diffusion LLM function structure.](https://ss.rapidrecap.app/screens/Bqdf6Um_8OE/00-00-01.jpg)
![Screenshot at 01:34: 00:Visual demonstration of the Autoregressive LLM generating 'The quick brown fox jumps...' sequentially, highlighting the committed nature of each token.](https://ss.rapidrecap.app/screens/Bqdf6Um_8OE/00-01-34.jpg)
![Screenshot at 08:51: The voice-enabled assistant responding instantly to a factual query about store hours, showcasing low latency in real-time conversational use cases.](https://ss.rapidrecap.app/screens/Bqdf6Um_8OE/00-08-51.jpg)
![Screenshot at 10:03: The fs-explorer tool executing a RAG query, showing the 2-step process \(Semantic Search then Get Document\) leading to the final answer with citations.](https://ss.rapidrecap.app/screens/Bqdf6Um_8OE/00-10-03.jpg)
![Screenshot at 11:01: A quality benchmark table comparing Mercury 2 against several leading models across metrics like GPQA Diamond, SCICode, and Average Quality.](https://ss.rapidrecap.app/screens/Bqdf6Um_8OE/00-11-01.jpg)