Linear attention has spent years as the thing I’d reach for only after giving up on quality to buy throughput. The Kimi Linear tech report is interesting because it claims the opposite trade — matching or beating full attention while cutting the memory you pay for at serving time.

• 🎯 Kimi Delta Attention (KDA) extends Gated DeltaNet with finer-grained gating, squeezing more out of the finite-state RNN memory that linear attention lives inside.
• ⚡ Up to 75% less KV cache and ~6× decoding throughput at 1M context — the numbers that actually move an inference bill.
• 📊 3B activated / 48B total, a layerwise hybrid of KDA and Multi-Head Latent Attention that beat full MLA on every evaluated task under an identical training recipe.
• 🔍 The claim that matters for LLM ops is drop-in replacement, not a long-context-only special mode — it held up in short-context and RL-scaling regimes too.
• 💡 They open-sourced the KDA kernel and vLLM implementation, so the throughput number is checkable rather than a benchmark screenshot.

KV cache is the quiet line item that decides how many concurrent agent sessions a single GPU can hold. If a hybrid linear stack really gives back three-quarters of it without a quality tax, the math for long-horizon agentic workloads — where context just keeps growing — shifts more than any single benchmark suggests. The HN discussion is already poking at the “under fair comparisons” caveat, which is exactly the right thing to poke at. Does the drop-in claim survive contact with a real 1M-context agent loop, or does the delta rule start forgetting things full attention would have kept?