Newsroom
· 11 min read · Release
Open-sourcing lithos-metal: Generating Megakernels for Apple Silicon
How lithos-metal compiles Gated DeltaNet, full attention, and DSpark drafting into GPU task schedules for Apple silicon.
· 4 min read · Release
Why we’re launching LithosBox, ultra-fast sandboxes for low-latency agents
LithosBox is now in public preview with two weeks of free usage. Ultra-low-latency sandboxes give agents Git-like control over their running environments: snapshot progress, branch into parallel attempts, inspect changes, and restore an earlier state.
· 5 min read · News
Open Weight Models Are No Longer Just Catching Up
Kimi K3, released today by Moonshot AI, sets the frontier on sustained agentic work rather than chasing it, with open weights, a native 1M-token context window, and launch pricing 40% under the closed flagships.
· 5 min read · Article
The Need for Speed in Agentic Inference
For agents, inference runs inside a loop, so speed — not just model quality — decides how fast real work gets done. LithosAI serves a 1T-parameter model at over 1,000 tokens per second per user on standard GPUs, at full model quality.