AX Gemma 4 31B IT MLX OptiQ 4-bit Assistant MTP

Parameter count: approximately 31.27B logical target parameters (31B class). 4-bit is the target quantization precision, not a 4B model-size claim. The separately packaged assistant is an acceleration component and is not included in the target count.

This is a self-contained MLX package for Apple Silicon. It combines the pinned upstream OptiQ target with its paired BF16 Gemma assistant for AX Engine multi-token prediction (MTP) / speculative decoding.

Try this model locally with AX Engine.

Attribution and changes

AutomatosX did not train Gemma 4, create the OptiQ quantization, or train the assistant. The target comes from mlx-community/gemma-4-31B-it-OptiQ-4bit at revision 23616162c5a8f928cac5b21d3e974d1dbc0b9877. The assistant comes from mlx-community/gemma-4-31B-it-assistant-bf16 at revision 28e92270316e89288579ec59c17939541d9ca433.

AutomatosX packaged the pinned target and assistant together, generated native manifests, applied the target tokenizer to the assistant subtree so both sides use identical token IDs, and added the exact-pairing contract. The target and assistant weights themselves are unchanged. The upstream OptiQ card is preserved as UPSTREAM_README.md.

No new training or benchmark results are claimed by AutomatosX.

Package details

Download and serve

hf download AutomatosX/AX-Gemma-4-31B-IT-MLX-OptiQ-4bit-Assistant-MTP \
  --local-dir ./AX-Gemma-4-31B-IT-MLX-OptiQ-4bit-Assistant-MTP

ax-engine doctor \
  --mlx-model-artifacts-dir ./AX-Gemma-4-31B-IT-MLX-OptiQ-4bit-Assistant-MTP

ax-engine serve ./AX-Gemma-4-31B-IT-MLX-OptiQ-4bit-Assistant-MTP --port 31418

The download is approximately 23 GB. AX Engine reads ax_gemma4_assistant_mtp.json, loads the drafter from assistant/, and uses the target model to verify every accepted token. Do not use assistant/ by itself as a general-purpose chat model.

The server exposes an OpenAI-compatible API. Consult the AX Engine repository for installation and API examples.

Validation and provenance

Validated on macOS arm64 with AX Engine 6.9.0 on 2026-07-20:

• AX artifact doctor: ready, with no model issues
• Safetensors headers, data bounds, and index mappings: passed
• Target weight shards: byte-exact against the pinned OptiQ source
• Assistant weight: byte-exact against the pinned assistant source
• Pairing contract: exact

See ax_provenance.json for immutable source revisions, transformations, and artifact hashes.

Quantization can change model quality, and speculative-decoding acceptance depends on the workload. Evaluate this package on your own tasks.

License

Apache License 2.0. See LICENSE, the Gemma 4 license page, and the official Google model cards for limitations and responsible-use guidance.