Instructions to use PINKlab/OSHE-Local-v2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use PINKlab/OSHE-Local-v2 with MLX:
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # if on a CUDA device, also pip install mlx[cuda] # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("PINKlab/OSHE-Local-v2") prompt = "Once upon a time in" text = generate(model, tokenizer, prompt=prompt, verbose=True) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- MLX LM
How to use PINKlab/OSHE-Local-v2 with MLX LM:
Generate or start a chat session
# Install MLX LM uv tool install mlx-lm # Generate some text mlx_lm.generate --model "PINKlab/OSHE-Local-v2" --prompt "Once upon a time"
- Atomic Chat
OSHE Local v2 Community Preview
OSHE Local v2 is a free, local-first AI engine for Apple Silicon Macs, developed and published by byPINKLAND Limited. It combines a Gemma 4 E2B foundation checkpoint with automatic capability routing, source-grounded multi-document analysis, bounded local file operations, local memory and a calculator Harness behind one local API.
The release includes its admitted signed learned capability pack. Users do not download, install or select adapters separately.
Why OSHE
An unmodified local checkpoint provides text generation. OSHE adds the product layer needed to use that checkpoint as a controlled local engine:
- automatic routing between chat and structured capabilities;
- source-grounded document comparison with conflict and missing-evidence checks;
- a signed, verified multi-document capability pack;
- bounded local file and memory operations;
- deterministic calculator Harness support;
- an OpenAI-compatible interface on a loopback-only endpoint;
- signed install, in-place update and rollback workflows.
Measured route improvement
On the same frozen eight-case multi-document subset and with the same evaluation gates, the complete OSHE Local v2 route improved over the unmodified Gemma 4 E2B base route as follows:
| Measurement | Gemma 4 E2B base | OSHE Local v2 | Change |
|---|---|---|---|
| Reference-score attainment | 65% | 95% | +30 percentage points |
| Schema compliance | 87.5% | 100% | +12.5 percentage points |
| Non-overconfidence | 87.5% | 100% | +12.5 percentage points |
These are route-specific product measurements for multi-document analysis. They measure the complete OSHE route, including its signed learned capability pack and required source-grounding guard, against the unmodified base checkpoint on the same cases. They are not a broad model benchmark, a general-intelligence claim or a claim of parity with larger foundation models.
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. This
known limitation remains disclosed and has not been relabelled as a pass.
See EVALUATION.md for the evaluation scope, metric definitions, reproducibility hashes and limitations.
Download
Release: 2.0.0-preview.1
Platform: Apple Silicon macOS
Archive: OSHE-Local-v2-Community-Preview-2.0.0-preview.1-macos-arm64.tar.gz
SHA-256: 6f1f4cec77b225e8b6b96215d08e1b26e55cebb7edaaec0a0166174e7c63216e
Verify the checksum before extraction, then follow INSTALL.md.
Local API
Base URL http://127.0.0.1:8093/v1
Model oshe-local-v2
Chat /chat/completions
Automatic /oshe/respond
The automatic route selects a supported capability and fails closed when a request is ambiguous or outside the public capability contract.
Terms and privacy
OSHE Local v2 is binary Community Preview software, not an open-source release. Read the included Terms, Privacy Notice, Model Lineage, and Third-Party Notices before use.
Publisher: byPINKLAND Limited, company no. 13945554
Registered office: 71-75 Shelton Street, Covent Garden, London WC2H 9JQ,
United Kingdom
Contact: hello@bypinkland.com