Two weeks after the Jev model captured attention across the AI industry, Cloudflare has unveiled its own pair of decision models that the company says outperform Jev while offering open weight access for local deployment. The Clef family, announced Thursday, includes two versions—Clef and Clef-flash—designed to tackle the same bounded, structured questions as TypeSafe's Jev: yes/no queries, multiple choice problems, and ranking tasks. But Cloudflare maintains its models deliver superior accuracy and speed, alongside capabilities Jev can't match.
The technical architecture separating Clef from competitors centers on its large language model foundation. Clef relies on specially post-trained, frozen versions of Qwen3.8-27B for the standard model and Qwen3.5-9B for Clef-flash, with the Qwen foundation executing a prefill-only pass during inference before scoring choices in parallel. In benchmark testing using the Jev Decision Index on Hugging Face, Cloudflare's internal results suggest Clef trades slight speed for higher accuracy compared to other open decision models, while Clef-flash matches competitors' accuracy but runs considerably faster. Against TypeSafe's own benchmarks, Cloudflare claims Clef beat Jev in three of four categories, losing only on agent trace observability. Beyond text classification, Clef handles images and video—a feature Jev lacks entirely. Both models support a 64,000-token context window, though Jev limits individual questions to 32,000 tokens.
Cloudflare AI Platform group product manager Michelle Chen confirmed to The Register that while described as "open source" in the announcement, the training datasets remain private. Chen specified hardware requirements: Clef-flash needs at least 41 GB of VRAM on a GPU, while the full Clef model demands 85 GB of VRAM, assuming single concurrency and a 64,000-token context window. The company hosts both models on Workers AI, which Chen said enables faster performance by "tak[ing] advantage of our GPUs at the edge, leading to low network latency and faster decisions." However, hosted access comes at a premium—Clef costs $0.24 per million tokens, nearly six times Jev's $0.042 per million rate.
The pricing gap reflects a strategic bet that multimodal capability and open weight flexibility justify higher costs for certain use cases. For organizations unwilling to pay Cloudflare's token rates, both Clef models are available for download from Hugging Face under Apache-2.0 license terms, the same framework governing Qwen. The models feature full Jev API compatibility, positioning them as drop-in replacements without requiring architecture rebuilds. That design choice lowers switching costs for teams already invested in decision model infrastructure, letting them test Clef's claimed advantages either through local deployment or Cloudflare's edge hosting. The competitive dynamics hinge on whether accuracy gains and multimodal features offset the steeper per-token expense, particularly as TypeSafe's underlying Jev architecture remains undisclosed. Organizations will need to weigh infrastructure investment against operational costs when choosing between proprietary speed and open weight versatility. The decision reflects broader tensions in AI deployment between vendor lock-in and the flexibility to run models on owned hardware.