Moonshot’s Kimi K3 goes live on Modal with a custom DFlash speculator
charles_irl · x · 2026-07-28
Moonshot ships Kimi K3 with a custom DFlash speculator for faster inference
Modal says Kimi K3 is now live on its platform, and that Moonshot is the first open 3T-class model it has launched with.
- Modal built a custom DFlash speculator for K3’s new architecture.
- The goal is to run K3 losslessly faster at day-zero launch.
- The attached chart shows throughput gains from DFlash block sizes 8 and 16 versus no speculation across different interactivity levels.
The post frames this as a launch-time optimization for a highly capable open model, not just a benchmark claim.
Related event: Moonshot's 3T-class Kimi K3 Lands on Modal(3 posts)→
More from Infra
- Poster says the NVIDIA-linked release is still gated to approved Microsoft Foundry customers — davidmanheim · 2026-07-28
- 8× B300 self-hosting could serve 30B tokens a month and pay back in under 100 days — JosephJacks_ · 2026-07-28
- Render open-sources an MCP server for deploys, logs, metrics, and Postgres — ojus_render · 2026-07-28
- Anduril’s containerized data center can be deployed by two people in under 10 minutes — damianplayer · 2026-07-28
- How teams run agents that need Docker, browsers, databases, or a GPU — michaelmanleyhypley · 2026-07-28
- Reddit user asks whether LLM routers can cut agent token spend by 30% — No-Nefariousness-728 · 2026-07-28