10,500 tokens/SEC per request on Nvidia hardware (morphllm.com)

Morph Labs today announced morph-v3-fast, a code-editing model running on Nvidia GPUs that reaches 10,500+ tokens/sec per request — achieved without batching or other latency tricks. That’s 2.3× faster than their prior best, 5× faster than basic search-and-replace workflows, and ~175× faster than “frontier” models, enabling edits that complete in human-perceived “instant” time. Real-world benchmarks include single-file edits (1k–3k tokens) in ~500 ms (vs 2.5–7.5 s traditionally), multi-file refactors (10k+ tokens) in ~1,000 ms (vs 25+ s), and a reported Fortune 100 fintech case of a 15k-token multi-file refactor in under 400 ms. The speed gain comes from a speculative architecture that layers semantic, structural, and context speculation with GPU-level optimizations: fused transformer ops to avoid extra memory writes, dynamic attention patterns tuned for code structure, and custom kernels for Hopper/Blackwell GPUs. Practically, that moves many coding tasks below the ~500 ms “invisible” threshold, unlocking speculative editing, instant collaborative diffs, agent swarms making parallel edits, and interactive refactoring at scale. Morph teases next steps—15k+ tok/sec, sub-100 ms latency for common edits, and real-time batch operations across hundreds of files—pointing to fundamentally smoother AI-assisted development workflows.