{{ t('models.manager.section_label') }}

{{ t('models.manager.heading') }}

{{ t('models.manager.description') }}

{{ t('models.manager.local_models') }}
{{ t('models.manager.table.model') }}

{{ t('models.downloader.section_label') }}

{{ t('models.downloader.heading') }}

{{ t('models.downloader.description') }}

{{ t('models.downloader.download_section') }}
{{ t('models.search.section_label') }}
{{ t('models.search.recent_searches') }}
{{ t('models.search.results') }}
B
B
GB
GB
{{ t('models.browse.section_label') }}

{{ t('models.browse.loading') }}

{{ t('models.browse.load_prompt') }}

{{ t('models.browse.searching') }}

{{ t('models.browse.search_prompt') }}

{{ t('models.oq.section_label') }}

{{ t('models.oq.heading') }}

{{ t('models.oq.description') }}

{{ t('models.oq.form_section') }}
{{ t('models.oq.enhanced_label') }}

{{ t('models.oq.enhanced_help') }}

{{ t('models.oq.imatrix_reuse_cache') }}

{{ t('models.oq.imatrix_reuse_cache_help') }}

{{ t('models.oq.imatrix_strict') }}

{{ t('models.oq.imatrix_strict_help') }}

Text Only

Excludes vision encoder weights. Output is a text-only model (~2-3% smaller).

{{ t('models.oq.preserve_mtp') }}

{{ t('models.oq.preserve_mtp_help') }}

{{ t('models.oq.preserve_mtp_unavailable') }}

{{ t('models.oq.dtype_label') }}

{{ t('models.oq.dtype_help') }}

{{ t('models.oq.no_models') }}

oQ: oMLX Universal Dynamic Quantization

Quantization should not be exclusive to any particular inference server.

oQ produces standard mlx-lm models that work everywhere — oMLX, mlx-lm, LM Studio, and any app that supports MLX safetensors format.

No custom loader required.

oQ and oQe: MLX-native mixed precision

The quantizer streams tensors from safetensors, measures layer sensitivity, and writes a byte-budgeted mixed-precision checkpoint. Standard oQ focuses on layer-level sensitivity and model-aware protection rules. oQe keeps that same oQ plan and adds imatrix-weighted affine quantization.

What standard oQ does
Layer sensitivity is measured with calibration prompts by comparing quantized-layer output error against the float baseline.
Mixed precision allocates higher bits to sensitive layers and protected tensors while keeping the target bits-per-weight budget under a hard cap.
Architecture rules keep MoE routers in fp16, protect shared experts and output-critical tensors, leave vision/audio encoders unquantized, and preserve SSM state tensors.
What oQe adds

oQe uses the same oQ sensitivity planner, then adds an importance-matrix calibration step. This is explicitly borrowed from the llama.cpp imatrix idea: collect activation energy from representative prompts, then use that information so quantization spends less error on the input channels that matter most.

Path Bit allocation Affine quantization Best use
oQ Layer-sensitivity mixed precision Standard min/max affine per group Fast, deterministic conversion with strong baseline quality
oQe Same oQ plan imatrix-weighted clipping/search per group Better low-bit retention, especially for MoE and activation-skewed layers
How oQe imatrix works
1
Collect activation energy. oQe runs calibration samples through the model and records average input activation squared, E[x^2], for each quantized Linear input channel.
2
Track experts separately. For MoE SwitchLinear layers, oQe records one importance vector per expert using the router-selected expert indices, so inactive or rarely active experts are visible in the coverage report.
3
Adapt sample count. Calibration starts from the requested sample count and can continue up to the adaptive maximum when MoE expert coverage is not sufficient.
4
Weight the quantization error. During affine quantization, oQe chooses scale/bias candidates by minimizing sum(importance * (weight - dequantized_weight)^2), not plain unweighted MSE.

If a tensor has no matching imatrix entry, the default behavior is to fall back to standard oQ for that tensor. Enable strict coverage to fail instead.

{{ t('models.detail.loading') }}