Skip to content
NewA calmer, faster ZeroHour
StoryModel releaseinfo

Cloudflare releases Clef decision models and a Managed Defense harness

1 source · 2 articlesFirst reported Updated ()2 · 1 read
What's new

The 7 October Cloudflare blog adds a Managed Defense harness that runs Clef on Workers AI to score likely false positives only after deterministic evidence collection, then specialist agents and approved models including GPT-5.6 Cyber and Anthropic Mythos, after a single-agent prototype hallucinated unsupported claims.

Merged summary · grok-4.7 · rewritten as coverage arrivesSummaries by Vani AI

Cloudflare open-sourced Clef and Clef-flash decision models and later described using Clef to filter false positives in a Managed Defense agent harness.

Cloudflare’s Workers AI team released Clef, a 27B model post-trained from Qwen3.8-27B, and Clef-flash, a 9B model post-trained from Qwen3.5-9B, under Apache 2.0 on Hugging Face and Workers AI, plus a reinforcement-learning fine-tuning service that The Decoder says uses Replicate technology. Both accept text, JSON, images, or video plus a schema and use a joint schema head to return one probability per allowed option in a single forward pass without free-form generation; MarkTechPost describes the types as yes/no, choice, and score, while the Clef-flash card says noul, choice, and score. Hugging Face cards report Decision Index 0.2.1 scores of BFCL 98.5, ToolRet 69.2 nDCG@10, and API-Bank 91.9 for Clef, and BFCL 98.8 and API-Bank 93.1 for Clef-flash, which that card says beat the larger model on those two; Cloudflare’s blog says the models lead the Jev Decision Index and credits Clef with a vision encoder and 64k context versus Jev’s 32k text-only design, while The Decoder says both support a 64,000-token context and images and attributes Jev to TypeSafe AI. Reported latencies differ only by rounding: MarkTechPost lists 209.3 ms for Clef and 38.8 ms for Clef-flash versus Jev’s 524.1 ms, and The Decoder rounds those to about 209 ms, 39 ms, and just over 524 ms; an internal threat-intel test classified a domain in 2.2 seconds versus 4.7 seconds for gpt-oss-120b. MarkTechPost, calling the figures vendor-reported and unreplicated, says a Clef model led 7 of 10 shortlist benchmarks, including BANKING77 macro-F1 of 94.20 versus Jev’s 79.74, while Jev leads GPQA Diamond, MMLU-Pro, and BBH. On 7 October Cloudflare described a Managed Defense harness that collects evidence with deterministic code before inference, uses Clef on Workers AI to score likely false positives, then routes remaining alerts to specialist agents whose typed findings a synthesis agent cannot expand; deeper analysis uses approved OpenAI Daybreak and Anthropic models including GPT-5.6 Cyber and Mythos after a single-agent prototype hallucinated unsupported claims.

Key facts
  • Clef (27B, post-trained from Qwen3.8-27B) and Clef-flash (9B, from Qwen3.5-9B) were released under Apache 2.0 on Hugging Face and Workers AI.
  • Both return one probability per schema option in a single forward pass with no free-form text; MarkTechPost lists yes/no, choice, and score, while the Clef-flash card says noul, choice, and score.
  • Decision Index 0.2.1: Clef scored BFCL 98.5, ToolRet 69.2 nDCG@10, and API-Bank 91.9; Clef-flash scored BFCL 98.8 and API-Bank 93.1, which its card says beat the larger model.
  • Median latency is reported as 209.3 ms for Clef and 38.8 ms for Clef-flash versus 524.1 ms for Jev; The Decoder rounds those to about 209 ms, 39 ms, and just over 524 ms and attributes Jev to TypeSafe AI.
  • MarkTechPost, calling figures vendor-reported and unreplicated, says a Clef model led 7 of 10 shortlist benchmarks, including BANKING77 macro-F1 94.20 versus Jev 79.74, while Jev leads GPQA Diamond, MMLU-Pro, and BBH.
  • An internal threat-intel test classified a domain in 2.2 seconds with Clef versus 4.7 seconds for gpt-oss-120b; a customer RL fine-tuning service is offered, which The Decoder says uses Replicate.
  • A 7 October Managed Defense harness collects evidence deterministically, uses Clef to score likely false positives, then specialist agents; deeper analysis uses GPT-5.6 Cyber and Anthropic Mythos after a single-agent prototype hallucinated.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

Coverage

oldest first · one row per article · UTC
  1. · Sep 30, 2026 Hugging Face trending modelsFirst report

    Cloudflare/clef — new model trending #27 on Hugging Face

    Cloudflare released Clef, a 27B open multimodal model that answers schema-constrained decisions in one forward pass.

  2. · Sep 30, 2026 Hugging Face trending models

    Cloudflare/clef-flash — new model trending #29 on Hugging Face

    Cloudflare released Clef-Flash, a 9B multimodal decision model post-trained from Qwen3.5-9B.