Cloudflare Open-Sources Clef: A 27B Multimodal Model That Turns AI Decisions Into Probability Outputs

A Decision Model That Generates No Text

Cloudflare's open-source Clef takes a fundamentally different approach from mainstream large language models. It's a 27B-parameter multimodal model, but its goal isn't to generate fluent text — it's to make structured decisions.

When traditional LLMs handle decision-making tasks, they typically generate free-form text first, then extract structured results through parsing — a process riddled with uncertainty and formatting errors. Clef eliminates this pipeline entirely: it takes a "state" and a "schema of typed questions," then returns a probability value for each allowed option of each question in a single forward pass.

No free-text generation, no output parsing. That means you no longer need to write regex to extract the model's answer, or worry about the model going off-script and producing unusable prose.

Technical Design: Multimodal Input, Probabilistic Output

Inputs Can Be Anything

Clef's "state" input is highly flexible — it can be plain text, JSON, images, or even video. This multimodal capability allows it to handle decision scenarios that span different data types, such as combining surveillance footage with a JSON configuration to determine what action a system should take.

Output Is Structured Probabilities

The key lies in the output side. For every question defined in the schema and every valid option for that question, Clef produces a corresponding probability. This design reframes "decision-making" from a generative task into a classification and probability prediction task.

The benefits are straightforward:

• Reliability: Outputs always fall within the predefined set of options — no hallucinated or invalid answers.
• Efficiency: All questions are resolved in a single forward pass, eliminating the need for multi-turn conversations or chained calls.
• Interpretability: Probability values inherently provide confidence information, making it easy for downstream systems to set thresholds or implement risk controls.

For production environments that demand determinism and auditability, this "probability in, probability out" pattern is far more robust than letting a large model generate freely.

From a machine learning perspective, Clef's design fundamentally models decision-making as multi-label classification or conditional probability prediction, rather than autoregressive text generation. Traditional autoregressive language models perform softmax sampling over the entire vocabulary (typically tens of thousands to hundreds of thousands of tokens) at each time step — a process that inherently introduces randomness and an uncontrollable output space. Clef instead constrains the output space to the finite set of options predefined in the schema, computing normalized probabilities for each option — an approach much closer to the discriminative model philosophy of the BERT era. A single forward pass means the model only needs to compute the network weights once in full, rather than looping token-by-token like a generative model. This provides a structural advantage in latency and compute consumption, making it especially well-suited for high-frequency inference scenarios.

Ecosystem Compatibility: Integration with Jev and SystemOne

Clef's API is designed to be fully compatible with Jev and SystemOne. This compatibility means developers already using either of these systems can integrate Clef at relatively low cost, without rebuilding their entire decision pipeline.

For a model that just went open-source, providing ready-made ecosystem integration points is a pragmatic way to lower the adoption barrier. It signals that Clef isn't an isolated research artifact — it's a tool designed for real-world engineering integration.

Jev and SystemOne are frameworks for AI inference orchestration, offering standardized model call interfaces, task routing, and result-processing pipelines. In AI engineering practice, "compatibility with existing orchestration frameworks" is a critically underrated metric. Development teams typically build logging, monitoring, and rollback mechanisms around a given framework. If a new model requires rewriting that infrastructure, the migration cost can far exceed the cost of learning the model itself. By aligning at the API level with these two systems, Clef can be dropped in as a replacement for existing decision nodes without touching upstream or downstream glue code. This is also a common strategy for open-source models entering enterprise deployment: winning adoption through protocol compatibility rather than exclusive features.

Why This Category of "Decision Model" Deserves Attention

Clef represents a growing school of thought: not every AI task requires a general-purpose large model that can hold a conversation. For the vast number of structured decision scenarios — routing choices, content moderation verdicts, risk scoring, configuration decisions — what we actually need is a stable, fast, controllable-output specialized model.

The fact that it's being open-sourced by Cloudflare, a company known for edge computing and infrastructure, also hints at its intended application domain: in high-throughput, low-latency scenarios like network request processing and security policy enforcement, a model that can deliver structured decisions in a single forward pass has a natural advantage.

On the community side, Clef received 91 upvotes on Product Hunt and ranked #8, appearing in the open-source and artificial intelligence categories. The response isn't explosive, but for a technically specialized decision model with a high barrier to entry, that level of attention is appropriate.

Cloudflare's core business includes CDN, DDoS protection, Zero Trust network access, and the edge computing platform Workers. These scenarios process millions of network requests per second, each potentially triggering a security policy decision, traffic routing choice, or content filtering judgment. At that scale, calling a general-purpose LLM that requires multi-turn dialogue or chain-of-thought reasoning is impractical in terms of both latency and cost. A decision model with a single forward pass and constrained output space is a natural fit for the low-latency demands of edge computing. Clef's open-source release can also be read as Cloudflare opening up internal infrastructure tooling to the broader community — similar to its previous open-sourcing of Workers Runtime — reinforcing its standing in the developer ecosystem through technical influence.

Summary

Clef's core value isn't in its parameter count — it's in the engineering answer it offers to the question of "how AI should make decisions." By eliminating free-text generation and directly outputting structured probabilities, it makes decision-making tasks controllable, auditable, and integrable. For teams building AI systems that require deterministic outputs, this is an open-source option well worth trying.