mlx-lm 0.32.2 loads Muse Spark from its Hugging Face checkpoint

Version 0.32.2 of mlx-lm is a small release, and Muse Spark users need to know about it. Muse Spark now loads from its Hugging Face checkpoint, using the Hugging Face names. The release also fixes attention value scaling in MiMo-V2-Flash, and padded prompts in state-space models.

Slide text

Muse Spark Fixes Upgrading

mlx-lm 0.32.2 loads Muse Spark from its Hugging Face checkpoint

Also fixed: MiMo-V2-Flash attention value scaling, and padded prompts in state-space models.

Until this release, Muse Spark support used its own names for the model type, the config and the weights. Now they match the Hugging Face checkpoint. The model type drops the word open. Text settings move under text config, with the Hugging Face field names. The language model weights load from model dot language model, and the module is renamed to match.

The inputs follow the same templates. The image processor now expands each patch token that the chat template writes, where it used to look for an image placeholder. The tool-call parser takes any parameter name inside double quotes, even after another attribute, and trims whitespace from text values. And NVFP4 models load when the quantization settings sit directly in the model config.

processor("A <|patch|> and B <|patch|>.",
          images=images)

<atem:parameter type="text"
  name="max tokens">two words</atem:parameter>

{"max tokens": "two words"}

Next, two fixes to model outputs.

MiMo-V2-Flash configs can carry a scale for the attention values, and until this release that field was ignored. Now the scale is read from the config and applied right after the value projection, before the cache, in both sliding-window and full attention layers. A new test checks that scaling the values gives the same output as scaling the value projection weights. Configs without the field behave as before.

if self.value_scale is not None:
    values = values * self.value_scale

The second fix is in the shared state-space code. When a prompt was padded and the model already carried state from earlier tokens, the padded positions still decayed that state. Now the decay is multiplied by the mask, so padding leaves the state alone, and a right-padded prompt matches the unpadded one. Every model that passes a mask through that code gets the fix.

# mlx_lm/models/ssm.py, ssm_attn
dtA = dt * A.reshape(1, 1, -1)
if mask is not None:
    dtA = dtA * mask[..., None]

Now, upgrading.

There's one breaking change, and it only affects Muse Spark. Version 0.32.2 drops the muse spark open model type, module and tool parser, and the processor stops looking for an image placeholder. So a Muse Spark model converted with 0.32.1 won't load. Convert it again from the Hugging Face checkpoint, and point any imports or prompts at the new names. Everyone else has nothing to migrate.

from mlx_lm.models import muse_spark_open
"model_type": "muse_spark_open"
"tool_parser_type": "muse_spark_open"
processor("Describe <image>", images=[img])

from mlx_lm.models import muse_spark
"model_type": "muse_spark"
"tool_parser_type": "muse_spark"
processor("Describe <|patch|>", images=[img])

To upgrade, run pip install with the upgrade flag. The one reason to wait is a Muse Spark conversion made with 0.32.1 that you can't convert again yet. Everyone else gets the MiMo-V2-Flash value scale and correct padded prompts in state-space models. And mlx-lm 0.32.2 loads Muse Spark from its Hugging Face checkpoint.