Artemis-31B-v1.2-oQ4e

An MLX oQ4e quantization of TheDrummer/Artemis-31B-v1.2, created with oMLX for Apple Silicon. Artemis is TheDrummer's fine-tune of Google's Gemma 4 31B instruction-tuned model, focused on creative writing, character dialogue, and roleplay.

The original model supports thinking and non-thinking operation using the Gemma 4 chat template. The upstream card includes community feedback on writing, reasoning, character consistency, and tool calling. Those reports are subjective observations about Artemis, not evaluations of this quantized checkpoint. This release changes weight precision; it adds no new training.

Model overview

Property	This checkpoint
Original fine-tune	TheDrummer/Artemis-31B-v1.2
Base model	google/gemma-4-31B-it
Model family	Gemma 4, 31B class, dense
Architecture	Gemma4ForConditionalGeneration (gemma4)
Text backbone	gemma4_text, 60 layers
Configured text context	262,144 tokens
Modalities present	Text and vision; no audio encoder weights
Format	MLX safetensors, mixed precision oQ4e
Weight size	19.00 GB / 17.70 GiB, weights only
Platform	Apple Silicon with a compatible Gemma 4 MLX runtime

The model retains a vision configuration and vision weights. Image input requires a compatible multimodal runtime and processor; image understanding has not been independently tested for this quantization. The configured context limit is not a guarantee of available memory or validated long-context quality.

Conversion details

Property	Value
Source checkpoint	TheDrummer/Artemis-31B-v1.2
Cached source revision used for this card	9007907f60c48a4f701cbbddaaf6353cdd078dfc
Quantization	oQ4e: sensitivity-based precision allocation with activation importance calibration
Default precision	MLX affine 4-bit, group size 64
Per-module overrides	5-bit or 6-bit, group size 64, as recorded in config.json
Weight files	4 safetensors shards
Exact weight file size	19,000,366,322 bytes
Calibration dataset	oqe_code_multilingual
Calibration samples processed	128
Calibration sequence length	512 tokens
Calibration report	oq_imatrix_report.json

The source revision above identifies the locally cached upstream model card; the output files do not independently record the source commit used by the converter. The quantization settings and sizes were read from this checkpoint's files.

Actual memory use includes model weights, KV cache, temporary buffers, image processing, and runtime overhead. Longer contexts require additional memory. The disk size alone does not establish a minimum RAM requirement.

Calibration notes

The report records coverage_sufficient: true and collection_sufficient: true, with no skipped layer forwards. Importance matrices were applied to 410 entries; one entry, embed_vision.embedding_projection, is listed as missing, with no shape mismatches. The text backbone is dense, so expert-coverage metrics do not apply. These calibration flags describe the collection process and are not quality benchmarks or a validation of vision performance.

Using this checkpoint with oMLX

Place the complete model directory in your configured oMLX model root, for example:

~/.omlx/models/Artemis-31B-v1.2-oQ4e/

Refresh the model inventory and select Artemis-31B-v1.2-oQ4e. Use an oMLX installation with Gemma 4 support in its MLX-VLM runtime. The calibration report identifies mlx_vlm.models.gemma4.gemma4.Model as the model class used during conversion. The Kolibri-specific MLX-LM fork mentioned in the Kolibri model card is not an Artemis architecture requirement.

Keep tokenizer.json, tokenizer_config.json, processor_config.json, and chat_template.jinja alongside the weights. The supplied Gemma 4 chat template handles conversation formatting and thinking mode.

Sampling and thinking

The included generation_config.json provides these defaults:

Parameter	Value
temperature	1.0
top_p	0.95
top_k	64

The Artemis author advises adjusting sampling to the task and using conservative settings when needed. See the original usage notes for community sampling resources.

The included chat template defaults to enable_thinking=false. Set enable_thinking=true when applying the template to request thinking mode. How this option is exposed through an API depends on the serving runtime. Tool-call parsing likewise requires support in that runtime.

Validation and limitations

File structure, configuration, weight sizes, and the calibration report were checked when preparing this card. Generation, throughput, long-context behavior, vision quality, tool calling, and quality relative to the source model have not been independently measured for this oQ4e checkpoint. Quantization can change output quality. No upstream benchmark or user testimonial should be read as a measured result for this release.

Attribution and license

Artemis-31B-v1.2 was released by TheDrummer; the underlying Gemma 4 model was developed by Google DeepMind. This is a community quantization.

The Gemma 4 base model card lists Apache 2.0. The cached Artemis card does not specify a separate license. This card uses the base model's Apache 2.0 license metadata; conversion does not grant additional rights. Preserve applicable license and attribution notices when redistributing the model.

• Original Artemis model card
• Gemma 4 base model card
• oMLX
• MLX-VLM

Quantization-specific details in this card were checked on 2026-10-05.