Repository navigation
0.11.1 (2026-11-11)
This release adds a Cognee integration for Python and TypeScript, onboards Muse Glimmer 30B, and makes the gateway point to the right model when a model id is unknown or retired.
Highlights
- Cognee integration: new
sie-cognee(PyPI) and@superlinked/sie-cognee(npm) packages run Cognee's embeddings, extraction and decisions on SIE, with source evidence helpers and runnable examples inexamples/cognee. - Clearer model-not-found errors: the gateway suggests the closest visible model ids on
MODEL_NOT_FOUNDand names the replacement when a retired model id is requested. - Placement planner: new
sie_placementplanner, tuner and cluster appliers with versioned input and plan schemas.
New models
- meta-models:
meta-models/Muse-Glimmer-30B.
Features
- add Cognee integration and memory primitives (#915) (b85d083)
- cognee: add TypeScript integration and examples (#921) (7271948)
- examples: look up official numbers in the benchmark example (#912) (027153a)
- examples: run the published-benchmarks example against your own server (#889) (991540b)
- examples: test a new model against an official number (#898) (3d62c0a)
- examples: time eight more website pages one request at a time (#931) (e91c9a2)
- examples: time one request at a time beside the website's times (#923) (f11ec14)
- examples: time the /detect page one request at a time (#937) (58748cd)
- gateway: name the replacement when a retired model id is requested (#944) (70fe04b)
- gateway: suggest the closest visible model ids on MODEL_NOT_FOUND (#942) (a64c300)
- models: onboard meta-models/Muse-Glimmer-30B (#888) (e8aeb68)
- placement: add the sie_placement planner, tuner and cluster appliers (#914) (c2cf482)
- placement: add versioned planner input and plan schemas (#904) (ba4faf0)
- server: reuse decision adapters for Qwen and Gemma families (#894) (2c22434)
- server: share label classification across decision adapters (#901) (57b572e)
Bug Fixes
- deps: update rustix to 1.1.6 for RUSTSEC-2026-0486 (#945) (9442378)
- helm: match reused chart archives by the chart they declare (#933) (32b55d3)
- models: raise the Muse Glimmer default output budget to 16,384 tokens (#897) (06b10dc)
- preserve native extraction token usage (#900) (aab8376)
- sdk: parse nested native content token usage (#902) (ae50f8d)
- server: keep memory free for full-size LightOnOCR-3-4B pages (#891) (4b104c2)
- server: keep Muse channel text from forming think and tool-call tags (#935) (d55df1a)
- server: render a null KnowLine noul criterion as its default (#934) (2a6c63a)
- server: report SGLang window refusals as client errors (#896) (5f21c08)