Top AI Product

Moonshot AI dropped Kimi Linear and it hit the HackerNews front page the same day (145 points). It’s not a chatbot or an agent — it’s an open-weight language model, but the news is the plumbing underneath it: a new attention architecture built to make long context cheap.

What it actually is

Kimi Linear is a 48B model (3B active) with a 1M-token context window. The trick is Kimi Delta Attention (KDA), a linear attention mechanism with fine-grained channel-wise gating, stacked 3:1 against full attention layers. Result: up to 75% less KV cache and up to 6.3x faster decoding at a million tokens versus standard attention. Attention Residuals push scaling efficiency ~2.5x over the older K2.

Why it matters

This is the opposite bet from the 2.8T flagship K3. Instead of piling on parameters, Moonshot is attacking the cost of long context at the mechanism level. Weights, a Triton KDA kernel, and vLLM inference are all open — pull the checkpoint and run it locally, no API gatekeeping.

You Might Also Like

• A Single api String Exposed Cursors Secret Composer 2 Runs on Moonshot ais Kimi k2 5
• Moonshot ai Ships Kimi k3 2 5t Parameters 1m Context Zero Benchmarks Published
• Kimi cli Hits 9600 Github Stars as Moonshot ai Walks Into Claude Codes Territory
• White House Accuses Moonshot ai of Distilling Anthropics Fable to Build Kimi k3 3 4 Million Claude Exchanges one Week Before the Weights Open
• Kimi Agentenv Moonshot Open Sourced the Agent rl Training Layer Nobody Else Shares