LOBSTERS · LOBSTERS// 40d agoPRODUCT LAUNCH

Mercury diffusion coder models hit 1,109 tok/s

Inception Labs’ Mercury paper introduces diffusion-based coding LLMs (Mini and Small) that generate tokens in parallel and report 1,109 and 737 tokens/sec on H100 GPUs. The work claims up to 10x throughput gains versus speed-optimized autoregressive models while staying competitive on coding quality benchmarks and Copilot Arena.

// ANALYSIS

This is a serious attempt to break the autoregressive latency ceiling for coding assistants, and the speed-quality tradeoff looks compelling if independent real-world evals keep holding.

–The key technical bet is parallel denoising over discrete tokens, which attacks serial decode bottlenecks directly.
–Reported throughput numbers are large enough to materially change UX for autocomplete, agent loops, and iterative coding chat.
–Quality claims are strong but still benchmark-heavy, so production reliability across messy enterprise codebases is the next proof point.
–If diffusion LLM serving matures, incumbent “fast” autoregressive coding models could face real pricing and latency pressure.

// TAGS

mercury-coderllmai-codinginferenceresearch

DISCOVERED

40d ago

2026-03-03

PUBLISHED

46d ago

2026-02-25

RELEVANCE

9/ 10