Clef Flash
A lightweight open-source decision-making AI model released by Cloudflare, representing a compact version of Clef. It substantially reduces inference latency while retaining decision-making capability, making it suitable for workflows with stringent response-time requirements.
Timeline (last 90 days)
As a personal demonstration from a single source lacking systematic performance data, this demonstration is better viewed as a proof of concept rather than a rigorous benchmark.
A developer had the Clef Flash model play Google's Snake game in real time on an RTX 5080 GPU, with no training involved and no exploitation of game state or algorithms.
The developer only provided the model with simple instructions (informing it of the snake head position, food direction, boundaries, and other basic information), and the model made real-time decisions based on this.
The model needs to complete the full perception–decision–output cycle within a 135-millisecond window to keep pace with the game.
The Clef Flash model used in this demonstration has 9B parameters.
Clef has 27B parameters, and Clef Flash has 9B parameters.
Both Clef and Clef Flash have multimodal capabilities; requests can contain both images and text, with images shared across all questions and jointly scored with text.
In the official example, Clef Flash returns a noul value of 0.958 for the query of whether an image contains Ollama.
Ollama v0.35.1 supports Cloudflare's open-source decision models Clef and Clef Flash, invoked via the /v1/systemone interface.
All Facts (6)
Ollama默认提供的模型多为4-bit量化版本（Q4_0或Q4_K_M格式）
85%UnverifiedThe Clef Flash model used in this demonstration has 9B parameters.
50%UnverifiedOllama v0.35.1 supports Cloudflare's open-source decision models Clef and Clef Flash, invoked via the /v1/systemone interface.
50%UnverifiedClef has 27B parameters, and Clef Flash has 9B parameters.
50%UnverifiedBoth Clef and Clef Flash have multimodal capabilities; requests can contain both images and text, with images shared across all questions and jointly scored with text.
50%UnverifiedIn the official example, Clef Flash returns a noul value of 0.958 for the query of whether an image contains Ollama.
50%