Solo dev makes vLLM tokenization 120x faster with Elixir in under three days
AccBalanced · x · 2026-11-22
A retweet highlights that inference.net shipped "the fastest tokenizer in the world," built solo by @TheHarryET in under three days. The linked article explains the team has spent months investing in a custom inference request router built in Elixir, and this change makes vLLM tokenization 120x faster — first in a series of engineering posts.
More from Infra
- Analysts clash over Google's TPU split: bet on DeepMind/Gemini or become the Nvidia of TPUs — vitaliychiley · 2026-11-22
- Amazon nears $7.5B deal to buy AI startup Decart, nearly double its last valuation — rohanpaul_ai · 2026-11-22
- SurrealDB scans 10+ billion rows per second with a single SurrealQL query — mattturck · 2026-11-22
- Maryland's 1,300 MW coal plant can't retire until a transmission line Sierra Club opposes gets built — paulnovosad · 2026-11-22
- Agentic workloads aren't polite batch jobs: they fan out, stall, and reshape mid-run — AccBalanced · 2026-11-22
- Inkling 2 Mini hits 117 tok/s on a single 4090 via vLLM, author says ditch llama.cpp — MerePotato · 2026-11-22