Inception Launches Mercury 3 Diffusion LLM, Hitting 1,068 Tokens per Second
Inception has released Mercury 3, the newest version of its diffusion language model. Unlike conventional autoregressive LLMs that write token by token, Mercury 3 generates answers in parallel. Inception reports throughput of 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—an increasingly relevant proposition in 2026 as latency-sensitive agentic workloads become the norm.
TL;DR
- Size: Parameter count not disclosed. Inception says the model is more than 40x larger than Mercury 1. Context window is 1M tokens.
- Runs on: Inception-served NVIDIA GPUs (exact GPU model not disclosed). API only—no open weights or quantized builds.
- Performance: Leads 8 of 12 benchmarks versus GPT-6 Luna and Claude Haiku 5.5 (self-reported, medium settings).
- Best result: SWE-Bench Atlas at 51.08, roughly 18 points ahead of both rivals.
- Worst result: DeepSWE v1 at 46.31, 5.31 points behind Claude Haiku 5.5 (51.62).
What Is Mercury 3?
Mercury 3 is Inception's fourth Mercury release, following Mercury 1, Mercury 2, and Mercury 2.5. Mercury 1 launched in 2025. Mercury 2.5 added tunable reasoning and a 260K context window, according to Inception's launch announcement.
Why It Matters
The release highlights a broader shift in 2026: diffusion-based language models are moving from research curiosity to production-grade infrastructure. Parallel decoding offers a structural advantage over token-by-token generation for workloads where throughput directly determines cost and user experience—high-volume agent loops, real-time voice interfaces, and large-scale retrieval and search pipelines. The 1M-token context window further positions Mercury 3 for long-document and multi-step agentic tasks, where context retention has become a key differentiator among frontier models.
via MarkTechPost