Inception Labs Launches Mercury 3, a 1000+ tok/s Diffusion LLM Rivaling GPT-6 Luna

Inception Labs (founded by Stanford professor Stefano Ermon) officially released the diffusion-based large language model Mercury 3 on October 20, calling it the strongest diffusion LLM to date: roughly 3x the intelligence of Mercury 1 on standard benchmarks, matching or exceeding speed-optimized frontier models like GPT-6 Luna and Claude Haiku 5.5 in quality, while sustaining 1000+ token/s output and a 1M context window. Why it matters: the company says Mercury 3 has over 40x the parameters of Mercury 1 at the same speed, suggesting diffusion LLMs have unlocked a new scaling path—growing intelligence without proportional latency increases, in contrast to the inference bottleneck of traditional autoregressive models.

Confirmed

Official release with public blog details: 1068 tok/s, 1M context, input/output pricing of $0.1/$0.35 (described as roughly half the price of comparable models)
Quality matching or exceeding GPT-6 Luna and Claude Haiku 5.5 (per official benchmarks)
Total parameters over 40x Mercury 1's with unchanged speed; official materials show the evolution across five Mercury generations
Deployed in agent scenarios: code review, interactive education, and browser operation; early customers include TinyFish, Lorikeet, and Bauplan

Unconfirmed

Benchmark comparisons and the claimed 3x intelligence gain are all vendor-reported; independent third-party evaluations are still lacking

Why It Matters

If diffusion truly scales, the cost and experience advantages of high-throughput, low-latency LLMs in agent and real-time interaction scenarios will widen further—Mercury 3's early agent use cases are direct evidence.

2026-10-20 ~ 2026-10-20 · 9 related posts

Primary sources

Mercury 3 launches: 1,068 tok/s, 1M context, 50% off at $0.10/$0.35 per million tokens — StefanoErmon · 2026-10-20 00:42 (UTC+08:00)
Mercury 3 benchmarks: beats GPT-6 Luna and Haiku 5.5 at 1,068 tok/s — StefanoErmon · 2026-10-20 00:42 (UTC+08:00)
Inception argues diffusion scales: Mercury 3 is 40x bigger yet just as fast — StefanoErmon · 2026-10-20 00:42 (UTC+08:00)

Inception Labs releases Mercury 3, a diffusion LLM claiming 1,000+ tok/s at GPT-6-level quality — StefanoErmon · 2026-10-20
[source] Inception argues diffusion scales: Mercury 3 is 40x bigger yet just as fast — StefanoErmon · 2026-10-20
[source] Mercury 3 benchmarks: beats GPT-6 Luna and Haiku 5.5 at 1,068 tok/s — StefanoErmon · 2026-10-20
[source] Mercury 3 launches: 1,068 tok/s, 1M context, 50% off at $0.10/$0.35 per million tokens — StefanoErmon · 2026-10-20
Inception Labs' Mercury 3: diffusion LLM hits 1000 tokens/sec, matching Luna/Haiku — Ok_Knowledge_8259 · 2026-10-20
Inception Labs ships Mercury 3: diffusion LLM beats GPT-6 Luna at 1,068 tokens/sec — volokuleshov · 2026-10-20
Mercury 3 shows diffusion LLMs scale: 40x bigger, 3x smarter, still 1000 tok/s — volokuleshov · 2026-10-20
Mercury 3 launches as the most intelligent diffusion LLM, outpacing GPT-6 Luna at 1000+ tok/s — chrispiech · 2026-10-20

1 near-duplicate retellings: sedielem