Clef-text-0.6B
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity6/10
Open PyTorch and Core ML weights, a detailed model card and a Swift runtime in FluidUse. No hosted API or SLA. Maturity reflects an accessible local model with one integrated runtime.
- Capability7/10
Choice, score and boolean in one pass. The author reports 79.1% across 12 held-out suites and ECE 0.046; these are not independently reproduced. Latency is a demo average, not p50/p99.
- Adoption2/100
At first check: 2 HF likes and no downloads; launch post had 75 likes and about 2.8K views. FluidUse has 301 stars, but those are project-wide, not model-specific. No external production case.
Vendor claims
- 79.1% gold accuracy over 12 held-out task suites (5,124 questions), compared with 88.2% for the Clef-flash teacher[Vendor claim — not independently verified]
- M5 Pro: about 48 ms per three-question support ticket on GPU and 77 ms on Neural Engine[Vendor claim — not independently verified]
- The Core ML port changed 15 of 5,118 answers; the author attributes them to near ties[Vendor claim — not independently verified]
News covering this model
Clef-text-0.6B is an open text decision model from FluidInference, published on 10 October 2026. The Hugging Face card calls it a 0.70B-parameter model: a Qwen3-0.6B backbone with Clef’s joint schema head, distilled from Cloudflare’s 9B Clef-flash. The card publishes both PyTorch weights and a Core ML package. The tweet rounds the model size to 0.6B.
It accepts a state and typed noul, choice or score questions and returns a probability for each allowed answer in one pass. The weights are Apache-2.0. The Core ML package exposes four input-length buckets, up to 16 questions and 96 options; the Swift ClefTextManager runtime is in FluidUse.
Evaluation claims
The author’s model card reports 79.1% gold accuracy over 12 held-out task suites (5,124 questions), versus 88.2% for the 9B teacher. It reports 86.4% teacher agreement and dev ECE 0.046. These are UnverifiedClaim: no independent rerun is linked, and the task mix includes knowledge and reasoning sets where the student trails the teacher substantially.
The card reports about 48 ms for a three-question support-ticket example on an M5 Pro GPU and 77 ms on its Neural Engine. Those are one demo’s timings, not a latency distribution. Its Core ML port differed from the PyTorch model on 15 of 5,118 answers (0.3%); the author says all cases were near ties. These figures are also UnverifiedClaim.
Why it is in the catalog
This is a published checkpoint trained specifically for typed decisions, with a documented question schema and a runtime that consumes it. It meets the catalog criteria as an open model. The scores reflect the published evidence but remain provisional until an independent evaluation.