Fetching from the wire…
01
02
03
04
05
06
07
08
Infra2026-08-06 · source-backed
Published August 4, Exponent Aligned ANS splits BF16 into its 8-bit exponent field and its sign/mantissa, aligns exponents to a block-local center, then entropy-codes only the low-entropy exponent stream while storing sign and mantissa raw. 1.51x lossless, 622 GB/s median decode (peak 628), 391 GB/s encode. Because decode far exceeds typical ~50 GB/s network bandwidth, remote KV cache transfer becomes bandwidth-bound rather than codec-bound, which cuts time-to-first-token. Shipping in OpenLake v0.8 (Apache-2.0, Rust, 2,307 stars) with vLLM integration.
Each link below shares sources, entities, or timing with this story.
Shared entities / Shared topic / Earlier coverage
Both cover Apache, H100, OpenLake, Rust; overlapping topics (apache-2, cache); earlier Apache coverage from 2026-07-26.
Zerostack uses Rust / Shared entity: Rust / Earlier coverage
Linked by a graph relationship (Zerostack uses Rust); both cover Rust; earlier Rust coverage from 2026-05-17.
Codex CLI uses Rust / Shared entity: Apache / Earlier coverage / Tension
Linked by a graph relationship (Codex CLI uses Rust); both cover Apache; earlier Apache coverage from 2026-07-28.
Mnemo built by Rust / Shared entity: Rust / Earlier coverage / Tension
Linked by a graph relationship (Mnemo built by Rust); both cover Rust; earlier Rust coverage from 2026-06-05.
Shared entities / Earlier coverage / Tension
Both cover Apache, H100; earlier Apache coverage from 2026-06-10; pushes against this story (competes).
Both cover Apache, H100; earlier Apache coverage from 2026-04-01; pushes against this story (vs).
Codex CLI uses Rust / Shared entity: Apache / Earlier coverage
Linked by a graph relationship (Codex CLI uses Rust); both cover Apache; earlier Apache coverage from 2026-08-05.
Grok uses Rust / Shared entity: Rust / Earlier coverage
Linked by a graph relationship (Grok uses Rust); both cover Rust; earlier Rust coverage from 2026-07-29.