On September 8, Inception Labs released Mercury 2.5, described officially as "the strongest diffusion-based large language model on the market." The most eye-catching number was speed: 1107 tokens per second.
Placed alongside two earlier data points, the number takes on a different meaning. In February 2025, the original Mercury ran at 1109 on H100; in February 2026, Mercury 2 moved to Blackwell and hit 1009—slightly slower than the first version; in September 2026, version 2.5 posts 1107. Nineteen months, three product generations, two hardware cycles, and speed has drawn an almost perfect circle back to where it started.
The Speed Dividend Went Elsewhere
The 2.5 upgrade list is long: context doubled from 128K to 260K, max output of 65536, adjustable reasoning tiers, parallel tool calling, and schema-based JSON output. The official intelligence claim is "40% smarter than Mercury 2"—with no named benchmark and no stated metric. This is the most inverted aspect of the launch: the first product once published a complete six-by-seven benchmark table, while the third generation names no benchmarks at all, comparing itself only to lightweight models in the same price tier—GPT-5.6 Luna Low, Gemini 3.5 Flash-Lite, Claude Haiku 4.5. Same price, different capability class.
A highly upvoted Hacker News criticism sums it up: speed is compared against lightweight models from two or three generations ago; intelligence is compared only against its own predecessor. Harsh, but accurate.
How Diffusion Differs from Autoregression
- Autoregressive decoding: emit one token, reread the full context, emit the next—serial, with steps equal to length.
- Diffusion decoding: fill the whole span with placeholder noise, predict all positions in parallel, iteratively denoise and refine, then converge. Parallel, with steps decoupled from length.
diffusing=true API parameter that streams intermediate denoising steps—essentially broadcasting the model's revision process, something autoregressive models cannot do.Independent Benchmarks
Independent data remains sparse. Artificial Analysis has not yet added 2.5; its measurements of Mercury 2 showed 657–933 tok/s, below the official 1009, with an intelligence index of just 12. A third-party AI BENCHY snapshot on September 2 recorded a 47% pass rate—strong on puzzle tasks, weak on tool calling, 1.31s average response. HN users who tried it called it "usable but far from frontier, roughly on par with the previous generation of open weights." Agentic coding is a recognized weakness; some users report it performs worse with custom execution frameworks.
Business and Team
Pricing: $0.20 per million input tokens, $0.75 output; launch discount of $0.04 and $0.15. OpenAI-compatible API, available on Baseten and OpenRouter. The most concrete customer testimonial comes from Augment Code: latency on one workflow cut from 150 seconds to 27, with costs down 90%.
The team is genuinely strong: CEO Stefano Ermon, Stanford professor and a founding figure in diffusion models, with co-founder Aditya Grover. A $50M seed round was led by Menlo Ventures, with NVIDIA, Microsoft M12, Andrew Ng, and Karpathy among investors. One widely circulated error worth correcting: Tri Dao is not part of the founding team; official materials never name him.
The launch also previewed two products: Mercury Voice, a speech diffusion model with under 170ms first-packet latency, and Mercury Router, which uses a diffusion model itself for model routing. Both point to the same judgment—diffusion's home turf is latency-sensitive, real-time scenarios.
A Narrowing Niche
As of September 2026, the diffusion landscape: Ant's LLaDA 2.1 has open-sourced ten-billion-scale weights, Dream 7B stays in academia, and DeepMind's Gemini Diffusion is the most capable but never sold. Mercury is the only commercially viable diffusion LLM that buyers can actually pay for and run.
Its pitch has shrunk from "5–10x faster" to four words: fast enough, cheap, low latency, 260K long context. The positioning is clear—voice agents, IDE inline completion, massive cheap subagents. Head-on competition with frontier autoregressive models is no longer on the agenda. Two observation points will judge this route: when independent benchmarks include 2.5, and where tool calling ranks on the official roadmap. The speed circle has been completed; the next one has to be drawn somewhere else.
---
Sources: Inception Labs' "Introducing Mercury 2.5" blog and pricing page; arXiv 2506.17298 (Mercury technical report); TechCrunch seed-round coverage; Artificial Analysis Mercury 2 measurements; AI BENCHY September 2 snapshot; Hacker News thread (2026-09-08).