Model Details
Clef-Flash
Quick answer: Clef-Flash is the smaller, latency-optimised member of Cloudflare's Clef decision-model family, released October 1, 2026. It is a 9B multimodal model post-trained from Qwen3.5-9B that returns calibrated probabilities for typed questions in a single forward pass. It scores 57.07 on the Decision Index (self-reported) at about 39 ms median and 122 ms p95 latency, and is released under Apache 2.0.
At a Glance
What it is: A 9B decision model that trades a few points of accuracy for roughly 5x lower median latency than Clef (38.8 ms vs 209.3 ms).
Where it leads: Best p95 latency (122.4 ms) among the compared decision models, top BFCL case-exact accuracy (98.8) and API-Bank (93.1), and the strongest result on Cloudflare's home-appliance simulator (97.7).
Where it lags: CLINC150+OOS macro-F1 (66.8 vs Clef's 97.4), RAGTruth hallucination F1 (35.6) and GSM8K (67.3).
What Clef-Flash Is
Clef-Flash uses Qwen3.5-9B as a frozen backbone with a joint schema head and rank-256 adapters, trained like Clef with label-smoothed cross-entropy, Brier-loss calibration and RLCD. It reads text, JSON, images or video and is API-compatible with Jev/SystemOne. On Cloudflare's four Typesafe workflow evals it beats Jev on customer service (77.0 vs 76.0) and is within a few points on the others.
Specifications
| Field | Value |
|---|---|
| Organization | Cloudflare |
| Hugging Face | Cloudflare/clef-flash |
| Base model | Qwen/Qwen3.5-9B |
| Parameters | 9B |
| Output | Typed probabilities (no text generation) |
| License | Apache 2.0 |
| Release date | 2026-10-01 |
| Availability | Hugging Face weights; Cloudflare Workers AI |
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| Decision Index (v0.2.1, chance-corrected) | 57.07 | Cloudflare decision leaderboard | 2026-10 |
Self-reported by Cloudflare on 36 of 38 Decision Index benchmarks (missing benchmarks scored as 0) and not yet reproduced by the upstream board. Per-benchmark results in the model card use a decision (option-probability) format and are not comparable with Benchgen's generative MMLU, GPQA or BFCL pages, so they are not added.