thaw is the fork() primitive for live LLM inference. snapshot a running session — weights, KV cache, and scheduler state — into a portable file, then fork N children that skip the cold prefill and diverge from the fork point. on an H100, a pre-warmed pool forks at 0.88s median per round versus a ~340s cold boot, bit-identical at the fork boundary. open source, pip-installable, works with vLLM and SGLang. we contribute upstream (vLLM fork/snapshot RFC #34303) and publish our research (re-feeding is not replaying, arXiv:2606.15621). built for RL post-training, parallel coding agents, and multi-agent systems — anywhere you fork from a shared context and pay the prefill tax on every branch.
| Website | https://thaw.sh |
| Employees | 5 (1 on RocketReach) |
| Founded | 2026 |
| Industry | Software Development |
Looking for a particular thaw employee's phone or email?
Matthew Yu is the Co-Founder of thaw.
1 people are employed at thaw.
thaw is based in San Jose, California.