Inception has released Mercury 3, the newest version of its diffusion language model. Mercury 3 is a diffusion LLM from Inception that writes answers in parallel, not token by token. Inception reports 1,068 tokens per second on NVIDIA GPUs and a 1M-token context window. On Inception’s own benchmark table, it leads GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests. For developers building agents, voice bots and search tools, this is a speed-first model with competitive scores.
TL;DR
- Size: Parameters not disclosed. Inception says it is more than 40x larger than Mercury 1. Context is 1M tokens.
- Runs on: Inception-served NVIDIA GPUs (exact GPU not disclosed). API only, no open weights or quantized builds.
- Performance: Leads 8 of 12 benchmarks vs GPT-6 Luna and Haiku 5.5 (self-reported, medium settings).
- Best: SWE-Bench Atlas 51.08, about 18 points ahead of both rivals.
- Worst: DeepSWE v1 46.31, 5.31 points behind Haiku 5.5 (51.62).
What is Mercury 3?
Mercury 3 is Inception’s fourth Mercury release, after Mercury 1, Mercury 2 and Mercury 2.5. Mercury 1 launched in 2025. Mercury 2.5 added tunable reasoning and a 260K context window, per Inception’s launch announcement.
How does a diffusion LLM work?
Autoregressive models write one token after another. Each new token waits for the previous one. A diffusion LLM starts from a rough draft of the full answer. It then refines every position over a few passes in parallel.
Fewer sequential steps means faster output on the same GPUs. The open question was whether this approach holds up as models grow larger.
Does diffusion scale?
Here Inception’s answer is yes. Mercury 3 has more than 40x the total parameters of Mercury 1 (mini). Average quality across standard benchmarks rose about 3x between those 2 models. Speed stayed above 1,000 tokens per second.
Inception team plots this as a quality vs size curve across 5 Mercury generations.
How does Mercury 3 perform on benchmarks?
All scores below come from Inception’s table. Every model ran at medium settings.
Mercury 3 posts its largest gains on instruction following and agentic coding. IFBench reaches 76.84, 15 points above GPT-6 Luna. SWE-Bench Atlas reaches 51.08, against 33.33 for Luna and 32.26 for Haiku 5.5. GPQA Diamond lands at 89.23.
The losses matter too. Haiku 5.5 wins DeepSWE v1 (51.62 vs 46.31) and MRCR v2 (78.40 vs 75.50). GPT-6 Luna edges ahead on AA-Omniscience-Public (46.94 vs 46.44) and AA-LCR v1.1 (78.30 vs 77.67). Long-context retrieval and longer coding tasks remain the weak spots.
