Fluid Inference releases 0.6B model running on Neural Engine, processes 1,000 tickets per minute

Fluid Inference has released clef-text-0.6b, a 0.70-billion-parameter text decision model built for Apple Silicon. The model was distilled from Cloudflare's 9-billion-parameter clef-flash into a Qwen3-0.6B backbone augmented with a joint schema head taken from the original model. The interface matches Clef and the Jev/SystemOne API: it accepts a state and batched questions (noul, choice, score) and returns a probability for every option on every question in a single forward pass. Weights are on Hugging Face at FluidInference/clef-text-0.6b; the Swift runtime is ClefTextManager inside the FluidUse package.

On official test splits (5,124 held-out questions) the model reaches gold accuracy within zero to four points of the teacher on routing and classification, topic, entity, sentiment, ranking, toxicity and emotion. On tasks that require world knowledge, science questions, commonsense QA and NLI, performance remains far below the 9B teacher. Teacher agreement stands at 86.4 percent, with an expected calibration error of 0.046 on the dev set. In short: a fast decision model, not a compressed encyclopedia.

A comparison of PyTorch and Core ML runs on the same 5,118 questions shows only 15 divergent answers (0.3 percent), all in near-tie cases where the top two options in PyTorch differed by no more than 0.11. Median probability deviation is 0.013; accuracy matches within ±0.7 points per task. A 512-token decode takes 35 milliseconds on the GPU, 64 milliseconds on the Neural Engine and 200 milliseconds on CPU alone. On the Mac tested the GPU is roughly twice as fast, but the Neural Engine leaves the GPU free for other workloads — a practical advantage in systems running graphics or parallel compute.

In the Swift environment a support ticket with three questions (about 370 tokens) is processed in 48 milliseconds on the GPU, the throughput that enabled 1,000 tickets in 57 seconds in the FluidUse demo. The same scenario takes roughly 77 milliseconds on the Neural Engine, versus about half a second for the original clef-flash (9B, Core ML on GPU). A 10× latency reduction makes the model a realistic candidate for a real-time router or classifier inside native Mac applications without sending requests to the cloud.

Distillation used 24,208 records from 12 public datasets (train splits; test splits held out) labeled by clef-flash in 8-bit Core ML (a 1-to-611 ratio versus fp32). The loss combined KL divergence at temperature 2 with 0.3 cross-entropy against gold labels. LoRA at rank 64 was applied to the backbone and the head was trained fully for two epochs, initialized from the clef-flash head weights where shapes matched. The model, the teacher and the Qwen3-0.6B base are all Apache-2.0 — open weights with full commercial freedom.