Mercury 3 Diffusion LLM Beats GPT-6 Luna: 8x Faster, 40x Scale, Agent-Ready

Mercury 3: A Diffusion LLM That Scales Without Losing Speed

On October 19, Inception Labs — founded by Stanford professor Stefano Ermon — released Mercury 3, a diffusion-based large language model that challenges the dominant autoregressive paradigm. The model's parameter count is over 40× larger than the first Mercury, yet it maintains a generation throughput of 1,068 tokens per second on NVIDIA GPUs, roughly 8× faster than GPT-6 Luna (~130 tokens/s) and 4.5× faster than Claude Haiku 5.5 (~240 tokens/s). Pricing during launch is $0.10 per million input tokens and $0.35 per million output tokens (standard pricing $0.20/$0.70). The context window expands to 1 million tokens, up from 260,000 in Mercury 2.5.

Benchmark Results: Beating Strong Autoregressive Models

Inception's disclosed benchmarks show Mercury 3 outperforming both GPT-6 Luna and Claude Haiku 5.5 across multiple evaluations:

GPQA Diamond (graduate-level science reasoning): Mercury 3 89.23 vs. GPT-6 Luna 86.36 vs. Claude Haiku 5.5 84.51.

SWE-Bench Atlas (real software engineering tasks): Mercury 3 51.08 vs. GPT-6 Luna 33.33 vs. Claude Haiku 5.5 32.26.

IFBench (instruction following): Mercury 3 76.84.

These results indicate that diffusion LLM is no longer just a speed demonstration; it can handle knowledge-intensive QA, complex instructions, and practical coding tasks at a competitive level.

Why Diffusion LLMs Are Fast: Parallel Generation vs. Sequential Prediction

Autoregressive models (GPT, Claude) generate tokens left-to-right, one at a time. This sequential dependency prevents full GPU parallelism. Diffusion language models instead predict many token positions simultaneously and refine the whole sequence over multiple denoising iterations. Each forward pass updates a batch of tokens, reducing the number of serial steps needed to produce a complete response. Even with several iterations, the overall wall-clock time can be lower than stepwise autoregressive decoding.

The article illustrates this with an analogy: autoregressive writing is like composing a sentence word by word; diffusion writing is like drafting a full paragraph and then polishing multiple parts in parallel.

Balancing Parallelism and Quality

Applying image diffusion to discrete language tokens introduces a trade-off: more parallel updates per iteration increase speed, but insufficient iterations can cause incoherent sentences, broken code, or flawed reasoning. Inception's technical reports describe a Transformer-based denoising network that operates from coarse to fine, and the team has iterated on training scale, denoising objectives, inference algorithms, and low-level kernel optimizations. Mercury 3 demonstrates that the speed advantage persists even as the model grows 40× and its comprehensive benchmark performance improves ~3× over Mercury 1 mini.

Agent-Era Role: Fast, Reliable Executor for Multi-Step Workflows

In AI agents, latency compounds across dozens of tool calls. The article presents two production case studies:

TinyFish browser agent : Integrated Mercury 3 into a web navigation system tested on 50 real-world Mind2Web tasks. Mercury 3 completed 25 tasks (50% success rate), notably higher than GPT-6.1 Sol, while using comparable token counts and steps at much lower per-token cost.

Octen AI real-time search for smart glasses : The search step takes ~80 ms; model generation was the bottleneck. Switching to Mercury 3 reduced median generation latency to 0.9 s and end-to-end response to 2.2 s, achieving the best speed-quality balance among tested high-speed models.

FlashAttention author Tri Dao commented that diffusion LLMs could serve as the ultra-fast "executor" paired with a smarter "coordinator/architect" model. This reframes model competition: not every agent step needs the most capable (and expensive) model; a division of labor between a planner and a rapid, cost-effective executor may be more practical.

Reference: Inception Labs blog — https://www.inceptionlabs.ai/blog/introducing-mercury-3 Tri Dao tweet — https://x.com/tri_dao/status/2112263206085554392