News

Small Models That Route Big Ones: Decision Models Arrive

ยท RenderBob team

Cloudflare, Amazon, Liquid AI, Microsoft, and Hugging Face released small open decision models in early October. They return labels and probabilities.

A tiny probabilistic router directs one request toward much larger language, vision, audio, or video specialists.

A cluster of unglamorous releases in early October points at a pattern: small models built to make decisions rather than generate frames or prose.

What was released

On 1 October, Cloudflare released Clef, its first in-house model, an open decision model on Workers AI priced at $0.24 per million input tokens and nothing for output. Amazon released Strands Decider 2B the same day, a 2-billion-parameter model for local CPU or GPU, with weights, training data, and scripts published. Liquid AI followed on 7 October with d1-3B and d1-omni-600M, open-weight models that return structured answers without generating output tokens. Hugging Face released JEV-27B-VL, a multimodal model that outputs calibrated probabilities from text and images. Microsoft released Microsoft-Decision-1, a fine-tuned Qwen3.5-9B for routing and classification. On 9 October Cloudflare added Clef-omni, an open-weight version that accepts audio, video, image, and text in one call.

What a decision model returns

A label, a score, or a choice from a list. That is fast, cheap, easy to run locally, and easy to audit.

Plausible uses in an edit

Many editing decisions are classification problems. These are plausible uses, offered as analysis rather than reported deployments:

  • Routing. Should this generation run on a local model or a cloud model, given its length, resolution, references, and deadline?
  • Shot classification. Is this clip a talking head, an establishing shot, or usable B-roll?
  • Quality screening. Did the generated clip drift from its reference? Is there garbled text or a warped logo?
  • Tagging and triage. Which of the 400 clips from this shoot belong in the rough cut?
  • Safety and policy checks. Does a clip break a client's content rules before anyone watches it?

A small model that screens outputs before a human sees them could cut the review load that comes with generative workflows. A small local model can do that without uploading footage.

Cautions

These are general-purpose decision models, not video-editing products, and early benchmarks come from their makers. Calibration matters. A model that is confidently wrong about a face match is worse than no model. Keep a human in the loop for anything a client will see.

What to watch

Fine-tuned decision models for film-specific tasks, and whether editors get a visible confidence score instead of a silent automatic choice.

More from the blog

All posts