v0.8.3

0.8.3 (2026-09-26)

Features

• examples: add a typed-decisions verification example (#362) (d4e00b3)
• examples: add watermark robustness eval (translation vs paraphrase) (#296) (a7c37ac)
• examples: refresh typed-decisions evidence on a faster server and add GLiNER2.5-Decide (#376) (8b4dc45)
• server: add GLiFormer multi-task extraction models (#358) (28a07f5)
• server: encode GLiClass label groups separately by default (#364) (71b395f)
• server: GLiClass task prompts, few-shot examples and label groups, plus instruct, Opir and multilingual models (#355) (bc1d66e)
• server: relations from GLiNER relex models, plus GLiNER bi-encoder v2 and PII models (#356) (5914e91)
• server: replay GLiClass forwards as CUDA graphs (#367) (7afc169)
• server: serve GLiNER2.5-Decide typed-decision models (#368) (2bbbc54)
• server: serve Laya typed-decision models through extract (#353) (5162aa7)

Bug Fixes

• examples: follow the screenshot-mining and multi-vector page selections (#349) (72c74bb)
• examples: score /structured-output's amended run, and drop the last display claim (#352) (26c81c2)
• examples: score the detect and image-classify pages as they now publish (#350) (e48c1a6)
• examples: score the redact composition, and stop both examples describing their pages (#357) (e9815c3)
• examples: stop chat and caption-vqa describing what their pages display (#351) (2e3c45e)
• examples: stop guardrails and speech-to-text describing what their pages display (#354) (79ae97e)
• gliformer: bound span decoding and keep padding out of it (#375) (0ee26b0)
• knowledge-graph: narrow two relation schemas and re-pin the evidence (#348) (72a9924)
• mcp: window GLiNER extract and redact by words so the whole text is read (#359) (35e711b)
• server: bound subword length of long words in GLiNER adapters (#380) (894a017)
• server: limit the label prompt of GLiNER-family extract requests (#381) (5aa0287)
• server: reject oversized item text at encode, score, and extract ingress (#366) (7bd6192)
• server: split GLiNER2 documents in linear time (#371) (359c4f0)
• sparse: score the three cards the page shows and the run behind them (#346) (c93f18a)

Performance Improvements

• gliformer: measure each task prompt once and skip the per-request eval() (#365) (d41ba7f)
• models: load three GLiClass models with bucketed CUDA graphs (#372) (a3a6429)