Reference-score attainment
Measured as a score-versus-teacher ratio on the fixed evaluation set.
Hugging Face
PINKlab local AI infrastructure
A free local AI engine for Apple Silicon, combining Gemma 4 E2B with automatic capability routing, source-grounded document analysis, local memory and an OpenAI-compatible integration layer.
Validated product performance
OSHE retains Gemma 4 E2B for general chat and applies its bundled signed specialist capability pack only to eligible multi-document tasks. The route combines learned task behaviour with deterministic source-grounding controls.
Measured as a score-versus-teacher ratio on the fixed evaluation set.
Responses that satisfied the structured multi-document response contract.
Responses that avoided overconfident conclusions when evidence was incomplete or conflicting.
Local text generation without OSHE routing, source checks, capability controls or integration contract.
The same local foundation for general chat, plus task routing, a signed specialist capability pack, source-grounded output and a local API.
Measured on the same frozen eight-case multi-document subset, scored as 32 records with the same evaluation gates and foundation-model hash: the unmodified Gemma 4 E2B base route versus the complete OSHE Local v2 route, including its signed learned capability pack and required source-grounding guard. These are route-specific product measurements, not a weights-only comparison, broad model benchmark or general-intelligence claim. The source evaluation reports all_gates_passed: false because a separate Cantonese mimicry diagnostic scored 0.125 against its 0.80 threshold; formal written Traditional Chinese passed 16/16 in the scoped product evaluation.
Product architecture
OSHE packages model serving, task routing, capability verification and host-side controls behind one consistent local interface.
General text generation is processed by the local foundation model without routing prompts to an OSHE-operated cloud service.
Compare supplied documents, identify conflicts and missing evidence, and maintain traceability between conclusions and source material.
The public multi-document specialist is installed with the engine, verified at launch and selected without manual adapter setup.
The /oshe/respond route selects chat, multi-document analysis, bounded local file access, local memory or the calculator Harness for each supported request.
Integration interface
Connect desktop applications, internal tools and reviewed bridges through the same OpenAI-compatible contract.
BASE_URL http://127.0.0.1:8093/v1
MODEL oshe-local-v2
CHAT /chat/completions
AUTO /oshe/respond
Deployment
Download the signed archive from Hugging Face. Before activation, the installer verifies the frozen inventory, release signature and required capability pack.
shasum -a 256 OSHE-Local-v2-Community-Preview-2.0.0-preview.1-macos-arm64.tar.gz
tar -xzf OSHE-Local-v2-Community-Preview-2.0.0-preview.1-macos-arm64.tar.gz
./bundle/bin/oshe-local install
6f1f4cec77b225e8b6b96215d08e1b26e55cebb7edaaec0a0166174e7c63216e
Community Preview
The Community Preview is provided for local evaluation, integration testing, technical review and feedback. It does not include service-level, support, performance, security or fitness-for-purpose commitments.
Included: engine runtime, foundation model files, public capability pack, installer and notices.
Publisher: byPINKLAND Limited, company no. 13945554.