Partners
TileRT runs frontier LLMs in real time.
GLM-5.3 on 8× AMD Instinct MI355X reaches 469 tok/s single-user generation on InferenceX AgentX, keeping ~2/3 of its short-context speed at 1M tokens.
Read the blog →Partners
TileRT, in motion.
Watch one Q&A unfold at real TileRT speed — no comparison, just the feel of it.
Single-user TPS on GLM-5 from 1K to 200K context. Holds steady where most engines collapse. More model results landing soon.
AI is going autonomous. In that world, what differentiates isn't intelligence — it's speed.
That's the gap TileRT is built for.
ChatGPT
Cursor · Claude Code · Codex
factories · agents · finance · vehicles
GLM-5.3 on 8× AMD Instinct MI355X reaches 469 tok/s single-user generation on InferenceX AgentX, keeping ~2/3 of its short-context speed at 1M tokens.
vLLM keeps prefill, scheduling, and the serving API; TileRT takes over decode at native single-stream speed — a zero-modification integration through vLLM V1's official connector interface.
Two execution-paradigm leaps to 1000+ tokens/s — a persistent execution model, microsecond-scale bottleneck triage, and model–system co-design with Xiaomi MiMo.
GLM-5.1-highspeed is now live on Z.ai, powered by TileRT — from experimental prototype to real production.
Inside TileRT and production-scale GLM-5.1 inference — persistent kernels, tile pipelines, and heterogeneous workers.
Support full size GLM-5-FP8 in TileRT with up to 500+ user decode TPS.
Multi-Token Prediction (MTP) enabled in TileRT, reaching up to 600+ user TPS for DeepSeek-V3.2.
Achieved 1.35x further speedup (3 ~ 4x speedup over baseline), reaching 250+ user decode TPS for DeepSeek-V3.2-Exp.
Initial public release, supporting DeepSeek-V3.2-Exp, achieving fastest inference speed among all available baselines.
TileRT is part of a growing ecosystem of tile-based AI computing.
Tile-based pythonic language for programming AI computing.
High-performance LLM operator library built on TileLang.
Distributed framework for AI computing across all scales.
Ultra-low-latency runtime for LLM inference.