v0.10.1

0.10.1 (2026-10-13)

Features

• add high-resolution Gemma thinking profile (#677) (e619232)
• api: opt-in reasoning_content on chat completions via include_reasoning (#728) (24400df)
• config: let a remote profile declare its own context window (#730) (530dcde)
• examples: add a full-frame Earnings21 URL caller (#680) (461189b)
• examples: add a portable Workflow DEV32 runner (#671) (5a6180f)
• examples: add image OCR and fixed-page evaluation (f7ba07f)
• examples: add portable Markdown frame runner (#670) (ca6980f)
• examples: coding agent with OpenCode, with an agent-loop conformance check (#727) (d59e8a1)
• examples: expose workflow32 top-k sampling (#686) (a23013c)
• examples: let the typed-decisions workflow32 runner set temperature (#679) (415f2f4)
• models: claim tool calling on GLM-5.3-Flash (#726) (3d39f37)
• sdk: default to SIE Serverless when an API key is set and no base URL is given (#691) (d680237)
• sdk: type SIE Relay refusals and attach request metadata to TypeScript errors (#689) (c571e0a)
• server: add LightOnOCR-3-4B model config (#666) (51f03ff)
• server: cap private reasoning with a per-profile thinking budget on SGLang (#700) (6c33b38)
• server: report SGLang vision extract prompt and completion tokens per item (#673) (b0020ff)
• server: serve LightOnOCR-3-4B on SGLang by default (#668) (a753964)
• server: serve the PII redaction composition as superlinked/pii-redact (#692) (766f799)
• server: show json_schema response formats to the model (#697) (f5eac2b)

Bug Fixes

• cluster: keep a worker lane up while its model is loading (#659) (f73cce5), closes #293
• examples: align document-ocr models with their image bundles (#654) (1fce5b5)
• examples: retain server timing headers in Earnings21 evidence (#690) (1689b17)
• examples: support the LightOn photo transcription output limit (2884f73)
• gateway: bound and re-estimate provisioning Retry-After, name the caller's model (#698) (ec068c4)
• gateway: name the model in cold-start refusals and let the transport set Retry-After (#688) (ce295a9)
• gliner2: time a Decide graph again instead of giving its shape up for good (#723) (46c736d)
• helm: keep the h200 machine profile in collector request telemetry (#702) (d0a76fa)
• models: keep json_schema out of the prompt on the Gemma 4 31B screenshot profiles (#706) (7ec754a)
• models: leave page-image headroom on the Gemma 4 31B answer-only hi-res profiles (#694) (c5b0d01)
• models: leave prefill headroom on the Gemma 4 31B hi-res thinking profile (#681) (7a22500)
• models: pin NuNER_Zero-span tokenizer dependency (#665) (c03a0c7)
• models: pin the tokenizer of gliner-relex-large-v1.0 (#715) (aff7ec4)
• models: put Qwen3Guard's stream interval on both profiles once (#687) (9fb03ab)
• replace unmaintained YAML parser in Rust services (#714) (5f6d451)
• sdk: bring TypeScript encode, score and extract options in line with the Python SDK (#704) (b76f698)
• sdk: declare the TypeScript SDK Apache-2.0 and read sie_sdk.version from package metadata (#705) (ab4c996)
• sdk: retry once more at the provisioning budget end and attach refusal metadata to capacity errors (#699) (a796d33)
• server: accept running usage in remote chat streams (#662) (d8053bd)
• server: accept tool_choice and parallel_tool_calls on the MLX chat path (#729) (a50706c)
• server: bound GLiFormer decoded text by characters (#655) (73f888e), closes #378
• server: keep upstream names out of remote profile error messages (#695) (c8c1c85)
• server: let Gemma 4 think before a grammar on the native generate path (#682) (0770521)
• server: let the IPC socket queue a burst of sidecar stream connects (#684) (f6e138d)
• server: release GLiClass CUDA graphs before evicting models (#657) (50a3817), closes #370
• sidecar: bound costly request waiting (#708) (8ec90e1)
• sidecar: bound model-readiness polling (#731) (c5abbe4)
• sidecar: keep the caller's key order of output_schema and options through to the engine (#724) (78fa1b4), closes #722

Performance Improvements

• gliner2: CUDA graphs and one pass per batch for GLiNER2.5-Decide (#721) (dfd25cf)
• laya: CUDA graphs for small requests, as an opt-in graphs profile (#718) (ab6c039)
• models: stream Qwen3Guard-Gen-4B every 16 decode steps (#683) (b1e3f3d)
• qwen3-reranker: fused kernels, shared prefixes and CUDA graphs over a cached prefix (#720) (e3576e7)
• server: give each SGLang OCR item its own worker batch (#675) (6dfce23)
• server: give GLiClass batches of up to eight rows a graph of their own size (#719) (2044b28)
• server: rebind the LightOnOCR-3 vision patch embedding (#669) (25a8353)
• server: serve more LightOnOCR-3-4B pages at once with SGLang overlap scheduling (#676) (b3546f8)
• server: use SGLang's fast image processor for GLM-OCR (#678) (2d2090d)
• sidecar: hold arrivals while the engine is busy and send them together when its batch returns (#725) (761a060)
• sidecar: send engine-batched extract items one per RunBatch, as wide as the adapter dispatches (#685) (4d13128)