Gemma-4-21B-REAP-JANG3M-MLX

Runtime Notice

This is an experimental JANG3M MLX package, not a fully quality-certified general-purpose release.

Validation snapshot, 2026-05-27:

• The published config.json and model.safetensors.index.json match the local oMLX copy reviewed by the maintainer.
• The model loads successfully with the maintainer's local oMLX / MLX runtime.
• A tiny generation smoke test completed without a shape-mismatch crash.
• Output quality is still experimental; use the included REAP/provider scaffold for guarded agent-style behavior.

If you see a shape-mismatch error while loading:

1. Make sure you are using a recent oMLX / MLX runtime with gemma4 support.
2. Re-download the repo into a clean folder instead of mixing files from older local experiments.
3. Confirm config.json, model.safetensors.index.json, and all shard files come from the same Hugging Face revision.
4. For stable baseline inference, prefer a validated oQ4 build until this JANG3M variant passes broader evals.

Summary

Gemma-4-21B-REAP-JANG3M-MLX is a local agent-instruct model package built around Gemma-4-21B REAP/JANG training work. It is intended for local agent-style use through a provider wrapper that can enforce structured-output behavior.

This is one model package, not separate model versions. The recommended Agent path uses:

• model weights
• tokenizer and chat template
• generation config
• included REAP runtime scaffold helper

Recommended Usage

Use the included examples / provider wrapper for Agent behavior. The included runtime scaffold helper can validate structured outputs, repair once, and return safe deterministic fallbacks for JSON/tool/5-tag/refusal routes.

No separate download is required for the scaffold; it is included in the package as helper code.

Technical Caveat

The scaffold is included in the repo, but .safetensors weights cannot execute Python logic by themselves. Some third-party runtimes may only load weights, tokenizer, and config, and may not automatically use the scaffold helper. For best Agent behavior, use the included REAP Agent examples/provider wrapper.

Known Limitations

• Strict JSON/tool behavior is not guaranteed by weights alone.
• Generic runtimes may not execute helper code automatically.
• Code generation remains weak in local evals.
• Runtime validation is a guardrail layer, not a proof of correctness.
• Small adapter experiments were tested locally during development; the final release keeps the clean JANG3M path.

Intended Audience

Local AI agent developers who want a compact REAP/JANG agent package with repo-included structured-output helper logic for local provider wrappers.