-
Card: install via `pip install typical-ai` (import typical_ai)
-
v2 card: backbone change to Qwen3.5-4B, corpus fix, and the CLINC/HWU64 regression
-
v2: replace weights with the fixed-corpus Qwen3.5-4B checkpoint (tm2)
-
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)
-
typical-medium (tm1b): weights, evals, inference package, manifest
-
Add inference/README.md (shared package doc, mentions all three releases)
-
Add README: product/launch tone, model family + lineage, compact results
-
initial commit