Skip to main content

Compression Jobs

Launch and track CompactifAI compression workflows for your models.

Compression Jobs is the home for CompactifAI: it lists every compression workflow that's been run or is running, and is where you launch new ones.

Before you start

The model you want to compress must already be in your project's Model HUB catalog. If it isn't there yet, import Qwen/Qwen3-0.6B from Hugging Face first, see Upload a Model from Hugging Face.

Recent Workflow list

ColumnWhat it shows
Recent WorkflowThe job's name, either auto-generated or one you set.
StatusSucceeded, Failed, or Cancelled.
CreatedRelative time since the job was submitted.
View / CancelView opens job details; Cancel is only enabled while a job is still running.

The CompactifAI Studio modal

Clicking New Compression Job opens CompactifAI Studio, where you configure a job before launching it:

FieldWhat it does
Model presetWhich model in Model HUB to compress.
Use CaseThe intended usage pattern (e.g. Agentic / Tool calling), used to tune the compression strategy.
Job nameOptional; auto-generated as {user_email}-profiler-{random_hex} if left blank. Must be unique.
Compression rateTarget size reduction, shown as a percentage; the UI marks a recommended "best range for reliable results."
Healing stepsWhich post-compression recovery stages to run: Recovery, Consolidation, Adaptation, and Localization. See Healing Steps below.
Compression speedGPU allocation for the job (e.g. Fast, 16 GPUs), which drives the estimated duration shown below it (e.g. 4-7 days).
Estimated costA live cost range for the job based on the settings above.

Healing Steps

After compression, the model needs to "heal": recover accuracy lost during pruning and, optionally, gain new capabilities. Healing runs as up to four sequential training stages, each continuing from the checkpoint of the one before. The Healing steps control picks how many of these stages to run; selecting a stage always includes every stage before it (shown as "Runs stages 1-N").

#StageWhat it doesDefault
1RecoveryRestores basic, coherent language generation. Pruning can leave a model unable to produce fluent text; this stage repairs that and re-establishes general language ability.Yes
2ConsolidationBroadens and stabilizes general knowledge and instruction-following over longer context, across general, math, and science data. Closes most of the accuracy gap left by compression.Yes
3AdaptationSpecializes the model for real-world use, refining instruction-following and task behavior.Yes (default)
4Localization (Spanish)Adds Spanish-language capability on top of Adaptation, producing a bilingual English/Spanish model. Additional specializations for this stage (such as tool calling) are planned for future releases.No (opt-in)

Which should I select?

  • The default (stages 1-3) covers general-purpose, English-language deployments.
  • Select all four stages if you need Spanish-language support.
  • Running fewer stages reduces training time and cost, but skips the capability gains described for the stages you leave out.

Click Compress to launch the job; it will appear at the top of the Recent Workflow list.