Card: install via `pip install typical-ai` (import typical_ai)

inference: Qwen3.5 support (VL unwrap, Gated-DeltaNet LoRA targets, linear_attn walk)

Card: document the long-document state-ordering caveat (measured, n=605)

Phase A serving speedup: zero-copy state-cache expand (no deepcopy per decision)

Serving-path speedup: state-KV LRU cache + suffix diet (see inference/bench_serve.py)

Serving-path speedup: state-KV LRU cache + suffix diet (see inference/bench_serve.py)

Serving-path speedup: state-KV LRU cache + suffix diet (see inference/bench_serve.py)

Serving-path speedup: state-KV LRU cache + suffix diet (see inference/bench_serve.py)

Fix latency claim, hard-tier SE claim, and make Quickstart copy-paste runnable

inference/README.md: mention typical-medium alongside small/preview

Rewrite README: product/launch tone, model family + lineage, compact results

How to run: document the public inference/ package

Add inference/: minimal self-contained Typical inference package

Publish release artefacts + model card + MANIFEST

initial commit