Fireworks event slides show sliding-window bit-sharing quantization trick
sophiamyang · x · 2026-11-04
sophiamyang retweets photos from Fireworks' "One way out" event, calling it amazing and mentioning meeting DHH and other smart people.
The slides explain a quantization idea where weights share a sliding window of bits: neighboring weights reuse bits, each new weight adds only 3 fresh bits, yielding 8 bits per weight with 512 possible values (vs 65,536 in EXL3), with weights chosen in pairs so rounding errors cancel out.
More from Infra
- Concurrency limits are the hidden tax in voice AI — how Phonic solved it — graceisford · 2026-11-04
- AMD beats guidance by just 2.9% as blowout era peaks; supply chain fungible between PCs and servers — tengyanAI · 2026-11-04
- Fireworks AI VP: Nexus Router Is Itself a Trained Model, Not Rules — sophiamyang · 2026-11-04
- Sipeed's 10cm MaixStation Dock claims DGX Spark-class local AI inference, runs off a power bank — vincent_koc · 2026-11-04
- vLLM, Red Hat and NVIDIA to host open-source inference meetup in Madrid on Nov 12 — vllm_project · 2026-11-04
- vLLM v0.33.0 ships with 927 commits, 1.5–2x faster decode on GB300 — vllm_project · 2026-11-04