FluidInference distills Clef-flash into a 0.70B on-device decision model
Source · Model card: Clef-text-0.6B
FluidInference published Clef-text-0.6B, a text decision model distilled from Cloudflare’s Clef-flash into Qwen3-0.6B with a joint schema head. The model card lists 0.70B parameters and Apache-2.0 weights; the launch post rounds the size to 0.6B.
The release includes PyTorch weights and a Core ML package intended for Apple’s Neural Engine. FluidUse adds a Swift ClefTextManager and a support-ticket demo. The post’s example triages 1,000 tickets on an M5 Pro in 57 seconds, assigning team, urgency and refund decisions. The card reports 79.1% accuracy across 12 held-out suites and about 48 ms for a three-question ticket on GPU, 77 ms on Neural Engine. These measurements are UnverifiedClaim; the release links no independent reproduction.
This is a trained checkpoint, so it belongs in the model catalog as well as the news timeline. It is a text-only student; Clef-flash remains the 9B teacher.