Cloudflare Introduces Multimodal Clef-omni, Reduces Clef-flash Pricing, and Speeds Up Clef

Cloudflare Introduces Multimodal Clef-omni, Reduces Clef-flash Pricing, and Speeds Up Clef Cloudflare has expanded its open-weight decision model lineup by launching Clef-omni with native audio and video support, lowering the price of Clef-flash by reducing its hosted context window, and accelerating Clef inference speeds through serving infrastructure improvements. Clef-omni adds audio and video processing to text and images without requiring separate transcription or captioning pipelines. Clef-flash received a price cut, making it cheaper than Jev, accompanied by a context window reduction on Workers AI from 64k to 24k tokens. Clef serving latency was improved by 1.7× to 2.0× at median through infrastructure optimizations, including migration to SGLang. Cloudflare teams internally deploy Clef for spam detection, plugin moderation, PII scanning, and malicious domain detection. Clef-omni Architecture and Performance Clef-omni processes text, images, audio (wav, mp3), and video (mp4, webm) directly in a single API call, removing the need for separate transcription or frame-splitting pipelines. Built on the Qwen3-Omni-30B-A3B-Instruct mixture-of-experts backbone without text-to-speech output generation, the model executes a single prefill pass across payloads and routes attention to score schema-constrained options. Median decision latencies reach around 130 ms for text, 150 ms for images, a few hundred milliseconds for audio, and roughly 1.5 seconds for a 21-second video with sound. Clef-flash Pricing and Context Adjustment Cloudflare reduced Clef-flash pricing below that of Jev. To achieve this lower price point, the hosted context window on Workers AI was reduced from 64k to 24k tokens, a threshold Cloudflare notes impacts only 0.24% of requests based on usage logs. The unhosted open weights on Hugging Face retain support for a 256k context window, while the standard Clef model remains hosted with a 64k context window. Serving Speedups and Internal Adoption Hosted Clef latency dropped by 1.7× to 2.0× at median across various token sizes following infrastructure optimizations, which included migrating model serving to SGLang (to be included in release 0.5.22). Internally, Cloudflare utilizes Clef across multiple operations, such as closing spam issues on GitHub documentation, moderating plugins in EmDash for phishing, scanning for PII in data loss prevention, and identifying malicious domains.

Cloudflare 发布全模态决策模型 Clef-omni，并优化 Clef 速度与 Clef-flash 价格 Cloudflare 宣布推出支持音频和视频的原生多模态开源决策模型 Clef-omni，同时通过服务架构优化将 Clef 的推理速度提升最高达 2 倍，并削减 Clef-flash 托管版上下文至 24k 以降低价格。 Clef-omni 基于 Qwen3-Omni-30B-A3B-Instruct MoE 架构构建，直接原生处理文本、图像、音频与视频并给出受约束的模式评分，不生成文本输出。 托管版 Clef-flash 价格已降至低于 Jev，权衡是将上下文窗口从 64k 缩减至 24k，但开源权重仍支持 256k。 Clef 迁移至 SGLang 推理引擎，800 至 16,000 token 下中位数推理速度提升 1.7 至 2.0 倍。 Cloudflare 内部已将 Clef 应用于 GitHub 文档垃圾 Issue 关闭、CMS 插件钓鱼审核、DLP 敏感信息扫描及恶意域名检测。 原生多模态决策模型 Clef-omni Clef-omni 改变了传统级联转写或拆分音画流的方式，允许在单次 API 调用中直接传入 wav/mp3 音频、mp4/webm 视频、图像与文本。模型基于 Qwen3-Omni-30B-A3B-Instruct MoE 骨干网络训练，冻结骨干并使用 LoRA，丢弃了文本转语音组件，因不作为生成式 LLM 运行而无需逐字解码。在延迟方面，文本决策中位数约 130 毫秒，图像约 150 毫秒，21 秒带声视频打分仅需约 1.5 秒。 Clef-flash 降价与上下文调整 Cloudflare 降低了 Clef-flash 的调用费用，使其定价低于 Jev。作为权衡，托管版本的上下文窗口从原定的 64k 压缩到 24k。数据显示仅有 0.24% 的调用请求会超过 24k token；若用户选择自建托管，Hugging Face 上的开源权重依然支持完整的 256k 上下文。 Clef 基础设施加速与内部实际应用 通过将推理后端迁移至 SGLang（支持 PR 已合并并将在 0.5.22 版本发布），Workers AI 托管的 Clef 获得 1.7 到 2.0 倍的中位数加速，约 3,400 token 处理中位数延迟从 616 毫秒降至 305 毫秒。Cloudflare 内部团队已开始使用 Clef 处理公共文档库垃圾 Issue、内容管理系统钓鱼检测、数据防泄漏（DLP）中的身份证件扫描以及威胁情报的恶意域名判定。

Following last week’s [release of Clef and Clef-flash](https://blog.cloudflare.com/clef-decision-models/), Cloudflare’s open-weight decision models, we decided to bring forth more gifts. Today, we’re releasing Clef-omni, which takes in audio and video input alongside text and image. We also cut the price of Clef-flash so it is now cheaper than Jev, and we made Clef faster. Although the model game is still early for Cloudflare, innovation and iteration is in our DNA, and we apply these principles to everything we do. In fact, the story of Clef came together over the course of less than a week. We decided we wanted to do something in the decision model space on a Friday evening, trained the model over the weekend, and launched it on Thursday. Even with such a short timeline, we were able to ship performant, high-quality, open-weight models for the community — imagine what more we can do in the future. For today, we’re excited to keep up the momentum with new additions and improvements to our Clef family of models. This is just the beginning, and we’ll continue to get better, faster, cheaper, and more innovative. Our new Clef-omni model is able to take audio, video, image, and text input. This changes the paradigm for decision models, which have been largely text-only since the debut of Jev from TypeSafe. With Clef, we supported images and video frame arrays, but Clef-omni is able to take in audio (wav or mp3) and video (mp4 or webm) alongside text and images. Instead of setting up cascading pipelines of models that transcribe speech-to-text, or splitting audio and image channels from video, you can just call one model to make decisions across any modality. We are now one step closer to a model that is able to interact with the world as we experience it — through audio, visual, and textual communication, all in one. Check out our developer docs for the new [Clef-omni](https://developers.cloudflare.com/workers-ai/models/clef-omni) model. We also released the model [open-weights on HuggingFace](https://huggingface.co/Cloudflare/clef-omni). For a quick start, here’s how you can send new modality inputs to Clef-omni: ``` curl https://api.cloudflare.com/client/v4/accounts/$CLOUDFLARE_ACCOUNT_ID/ai/run/@cf/cloudflare/clef-omni \ -X POST \ -H "Authorization: Bearer $CLOUDFLARE_AUTH_TOKEN" \ -H "Content-Type: application/json" \ -d '{ "model": "clef-omni", "state": "Review the installation: a photo of the unit, an audio recording of it running, and a video of the fan.", "images": ["data:image/png;base64,<base64-png>"], "audio": ["data:audio/mpeg;base64,<base64-mp3>"], "videos": ["data:video/mp4;base64,<base64-mp4>"], "questions": { "label_visible": {"type": "noul", "instructions": "Is the model and serial number label visible in the photo?"}, "sounds_normal": {"type": "noul", "instructions": "Does the unit sound like it is running smoothly, without rattling or grinding?"}, "fan_running": {"type": "noul", "instructions": "Is the fan running in the video?"} } }' ``` Clef-omni expands our open-weight decision architecture to natively handle multimodal workflows. We built this on a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts (MoE) foundation, where the model was already capable of processing text, imagery, audio, and video directly within a single pipeline. However, we adopt the primary comprehension backbone while discarding the text-to-speech output components. In production, Clef-omni executes a quick prefill pass across the complete payload, scoring all modalities and valid parameter options simultaneously. Because we skip output token generation since Clef models are not Large Language Models (LLMs), we remove the overhead of transcribing or captioning incoming files. Media elements map straight into the unified sequence, where video and audio are synced with visual frames for joint processing. Clef-omni then pulls candidate values directly from internal embeddings using our two-stage attention routing: every valid option gathers key evidence from the input (regardless if they are buried in text snippets, visual elements, or audio streams) before field vectors cross-attend across the full context to compute confidence scores. We add a built-in lexical grammar that preserves option semantics, so that we can deliver fast, schema-constrained scoring across every input type. We train the model with the same techniques as we did with Clef — freezing the Qwen3 backbone, training low-rank adapters (LoRA), and applying our standard post-training approach: combining label-smoothed cross-entropy loss with Brier score calibration. The result is a model that is resilient and designed to succeed regardless of schema variations, field ordering, and prompt structures. This brings several core upgrades to the Clef model family. Calibrated, schema-bound decisions now seamlessly handle text, images, audio, and synchronized video in a single API call. It's also fast. Text-only decisions return in about 130 ms at the median, image inputs in about 150 ms, and audio clips in a few hundred milliseconds. Even a full 21-second video clip with sound is scored in about 1.5 seconds, all in a single API call. Our benchmarking shows strong performance across benchmarks, even with new modalities and MoE architecture. | **Benchmark** | **Clef-omni** | **Clef** | **Clef-flash** | **Jev** | | BFCL · case exact | 98.2 | 98.47 | **98.76** | 95.75 | | ToolRet · nDCG@10 | 66.6 | **69.19** | 66.43 | 65.28 | | API-Bank · accuracy | 92.7 | 91.93 | **93.11** | 88.19 | | Home appliances · case exact | 69.3 | 82.95 | **97.73** | 52.27 | | When2Call · accuracy | 63.3 | 72.37 | 65.58 | **80.97** | | BANKING77 · macro-F1 | **94.8** | 94.20 | 90.93 | 79.74 | | CLINC150+OOS · macro-F1 | **97.7** | 97.43 | 66.77 | 89.27 | | BRIGHT · nDCG@10 | 42.0 | 45.91 | 39.26 | **47.52** | | Amazon ESCI · macro-F1 | **57.8** | 57.48 | 57.39 | 55.21 | | PhishNChips · accuracy | 73.2 | **79.60** | 75.05 | 62.55 | We also benchmarked against the [TypeSafe evals to show how Clef-omni performs](https://evals.typesafe.ai/). | **Workflow** | **Metric** | **Clef-Omni** | **Clef** | **Clef-flash** | **Jev** | |---|---|---|---|---|---| | Invoice processing | Exact actions | 60.2 | **64.7** | 57.1 | 61.8 | | Invoice processing | Primary action | 82.0 | **86.2** | 73.3 | 83.1 | | Customer service | Exact actions | 71.6 | 76.3 | **77.0** | 76.0 | | Security incidents | Exact actions | 61.7 | **62.9** | 61.7 | 61.7 | | Agent trace observability | Primary action | 65.8 | 68.5 | 69.8 | **71.6** | We heard your feedback — you want a decision model affordable enough to incorporate into any workflow. We’ve made a few optimizations and can now offer Clef-flash at cheaper prices. We want Clef to be able to make decisions that scale for you across all your agentic workflows, and now our pricing also incentivizes that. We were able to optimize the model to the point where Clef-flash is now *cheaper* than Jev while still being performant and high-quality. Check out the [developer docs](https://developers.cloudflare.com/ai/models/) for the most up-to-date pricing information at all times, including more detail on how image and audio modalities are converted to input tokens for pricing. However, one trade-off we have to make in order to have cheaper pricing is to cut the context window of Clef-flash. The hosted version of Clef-flash now has a context window of 24k, instead of 64k as previously advertised. The model weights on Hugging Face are untouched and were trained to support a 256k context window should you choose to self-host it. From our usage data, we see that only 0.24% of requests exceed 24k input tokens. Based on this data, we decided to make the call to drop the context window to 24k for Clef-flash in order to price it at a more accessible entry point. Our Clef model remains at a 64k context window, and we encourage folks with larger context needs to switch to Clef instead of Clef-flash. Speeds are also faster for the Clef model, with new optimizations to the hosted version of Clef on Workers AI. No new model weights are released because most of the optimizations we made are at the serving infrastructure layer, not materially to the model weights or architecture. Check out our new speeds below: | Input size | Before: median / p95 (ms) | Now: median / p95 (ms) | Median speedup | | \~800 tokens | 262 / 438 | 152 / 351 | 1.7× | | \~3,400 tokens | 616 / 777 | 305 / 531 | 2.0× | | \~16,000 tokens | 2,721 / 3,250 | 1,635 / 1,805 | 1.7× | One of the optimizations we made was to move to SGLang for serving the model. We worked with the SGLang team and made pull requests to incorporate Clef (PR [#42721](https://github.com/sgl-project/sglang/pull/42721)), which will be released in SGLang 0.5.22. You can also self-host Clef since we released the weights, and the new SGLang launch commands are available on our [updated Hugging Face repo](https://huggingface.co/collections/Cloudflare/clef) as well as the [SGLang cookbooks](https://docs.sglang.io/cookbook/autoregressive/Cloudflare/clef#hw=h200&variant=clef&quant=bf16&strategy=balanced&nodes=single). At Cloudflare, we’ve had teams experimenting with Clef as a model that can help classify over general domains. The coolest thing about watching teams adopt Clef is that the capability to do detections and classifications have moved into the model layer, where it is accessible for anyone to start incorporating decision models into their use cases. Previously, even with zero-shot classifiers, the small models may not have been powerful enough to classify across any domain without further tuning. You might have needed to stand up a specialized machine learning team to help train over your particular domain, and gather a corpus of data in order to train a better model. Now, all you need to do is call a model to instantly parse through input data with precision. A few highlights of how teams at Cloudflare are using Clef: Our public GitHub docs repo uses Clef to instantly [detect and close spam issues](https://github.com/cloudflare/cloudflare-docs/pull/34139). EmDash, our content management system, [moderates plugin libraries for phishing](https://emdashcms.com/blog/01M49N4779YCXX3QPKWSSNE3DY). Our data loss prevention team is scanning data for potential personally identifiable information (PII) like government IDs. And our threat intelligence team is using Clef to help detect malicious domains. We’re excited to bring you more from the Clef family of models, including new modalities like audio and video, as well as making the model more accessible with cheaper pricing and faster speeds. Clef is fully Jev-API compatible and available via AI Gateway, so it is as simple as changing the model ID to try out Clef today. Check out our [developer docs](https://developers.cloudflare.com/ai/models/) for more information. Thank you for all the love and warm reception around Cloudflare’s first models — we are focused on bringing you more. Try out our Clef family of models, and tell us what you think.