Mercury 2.5: Inception Labs' Third Diffusion LLM Launches at 1107 tok/s, But the Speed Story Hasn't Moved

On September 8, Inception Labs released Mercury 2.5, officially described as "the strongest diffusion-based large language model on the market." The launch's headline number was speed: 1107 tokens per second.

Set three data points side by side, and the number reads differently. In February 2025, the original Mercury ran at 1109 on H100; in February 2026, Mercury 2 moved to Blackwell and hit 1009—a bit slower than the first model; and in September 2026, 2.5 lands at 1107. Nineteen months, three product generations, two hardware generations, and speed has drawn an almost perfect circle in place.

Where Did the Speed Dividend Go?

The 2.5 upgrade list is long: context doubled from 128K to 260K tokens, 65,536 max output, 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 reference metric. This is the most inverted part of the launch: the original Mercury once published a full 6×7 benchmark table, while the third generation names no benchmark at all, instead benchmarking against same-price lightweight tiers—GPT-5.6 Luna Low, Gemini 3.5 Flash-Lite, Claude Haiku 4.5. Same price, different capability class.

A highly upvoted HN criticism put it bluntly: speed is compared only against lightweight models from two or three generations ago; intelligence is compared only against the previous in-house generation. Harsh, but the decomposition is accurate.

How Diffusion Differs from Autoregression

• Autoregressive decoding: emit one token → reread the full context → emit the next → loop until done (serial; steps scale with length)
• Diffusion decoding: fill the whole span with placeholder noise → predict all positions in parallel → iteratively denoise and refine → converge (parallel; step count is decoupled from length)

This structure gives diffusion models one form of structural advantage: fill-in-the-middle. Inserting a block of logic into the middle of code forces an autoregressive model to generate left-to-right pretending it's sequential, while diffusion is naturally bidirectional—spread it out and edit. The original Mercury scored 84.8 on the fill-in-the-middle benchmark, beating the then-strongest Codestral 2501 at 82.5—the hardest technical claim the diffusion route has made so far. Version 2.5 keeps dedicated Mercury Edit 2 endpoints for fill-in and Next-Edit, and the API exposes a diffusing=true parameter to stream intermediate denoising steps—essentially watching the model revise its draft live, which autoregressive models cannot do.

Independent Measurements

Independent data remains sparse. Artificial Analysis has not yet indexed 2.5; for the previous Mercury 2 it measured 657–933 tok/s (below the claimed 1009) with an intelligence index of just 12. Third-party snapshot AI BENCHY tested on September 2: 47% pass rate, strong on puzzle tasks, weak on tool calling, 1.31s average response. HN users who tried it said it's "usable, but far from frontier, roughly on par with the previous open-weight generation"; agentic coding is a recognized weak point, with one user reporting that adding a custom execution framework made it worse.

Business and Lineup

• Pricing: $0.20 per million input tokens, $0.75 output; launch discount 80% off ($0.04 / $0.15)
• Availability: OpenAI-compatible API; live on Baseten and OpenRouter
• Customer proof: Augment Code cut one pipeline's latency from 150s to 27s with 90% cost reduction

The team is genuinely stacked: CEO Stefano Ermon, Stanford professor and one of the founders of the diffusion model direction, with co-founder Aditya Grover. The $50M seed round was led by Menlo Ventures, with NVIDIA, Microsoft's M12, Andrew Ng, and Karpathy on the list. One widely circulated error to correct: 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 first-packet latency under 170ms, and Mercury Router, using diffusion models for model routing. Both point to the same judgment: diffusion's home turf is latency-sensitive, real-time scenarios.

A Narrowing Niche

The September 2026 diffusion LLM landscape: Ant's LLaDA 2.1 open-sourced tens-of-billions-scale weights, Dream 7B remains academic, and DeepMind's Gemini Diffusion is the most capable but never sold. The only commercially viable, well-funded diffusion LLM you can actually buy is Mercury.

Its pitch has shrunk from "5–10x faster" to four words: fast enough, cheap, low latency, 260K context. The positioning is clear—voice agents, IDE inline completion, massive cheap subagents; head-on competition with frontier autoregressive models is off the table. Two watchpoints for judging this route: when independent benchmarks index 2.5, and where tool calling ranks on the official roadmap. The speed circle has been drawn; the next one has to be drawn somewhere else.

---

Sources: Inception Labs official blog and pricing page ("Introducing Mercury 2.5"); 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).