Inference.net: ClipTagger 12B

This model is no longer available.

Add AI to your application with Puter.js.

Explore Other Models

Model Card

ClipTagger 12B is a 12-billion-parameter vision-language model built by Inference.net in collaboration with Grass, based on the Gemma-12B architecture and trained through knowledge distillation from a frontier VLM on a million curated video frames.

It is purpose-built for frame-by-frame video captioning at scale, returning a fixed JSON schema per frame with description, detected objects, actions, environment, content type, production quality, and logos.

Using Gemini-2.5-Pro as judge, it scored 3.53 on caption accuracy against GPT-4.1's 3.64 and Claude 4 Sonnet's 3.16, with ROUGE-1 of 67.4%, ROUGE-L of 52.0%, and BLEU of 26.7% against its teacher model. At roughly $335 per million generations, it costs about 15x less than GPT-4.1 and 17x less than Claude 4 Sonnet.

It suits teams building searchable video databases, content moderation pipelines, or ad verification and analytics tools, and is released under Apache-2.0.

Context Window 125K

tokens

Max Output 2K

tokens

Input Cost $0.3

per million tokens

Output Cost $0.5

per million tokens

Release Date N/A

Code Example

Add AI to your app with the Puter.js AI API — no API keys or setup required.

// npm install @heyputer/puter.js
import { puter } from '@heyputer/puter.js';

puter.ai.chat("Explain quantum computing in simple terms").then(response => {
    document.body.innerHTML = response.message.content;
});

<html>
<body>
    <script src="https://js.puter.com/v2/"></script>
    <script>
        puter.ai.chat("Explain quantum computing in simple terms").then(response => {
            document.body.innerHTML = response.message.content;
        });
    </script>
</body>
</html>

More AI Models From Inference.net

Find other Inference.net models →

Chat

Schematron V2 Small

Schematron V2 Small is a 3-billion-parameter HTML-to-JSON extraction model from Inference.net, fine-tuned from Meta's Llama 3.2 3B. It takes a JSON schema and raw HTML as input and returns schema-conformant structured output; extraction instructions go through the schema itself rather than a system or user prompt. Inference.net built it as the quality-focused tier of its V2 lineup, approaching 8B-level accuracy at roughly 3B-level speed, alongside a faster Turbo variant. On Inference.net's internal LLM-as-judge benchmark it scores 4.06 out of 5, close to the original Schematron 8B's 4.07 and ahead of the first-generation 3B's 3.91. On SimpleQA it scores 83.1, and Inference.net reports both V2 models beating DeepSeek V3.2 and GPT-5.4 Nano on that benchmark despite being much smaller. It targets web scraping, product data ingestion, financial document parsing, and search-augmented pipelines needing reliable structured output over long pages, with a 128K context window and 4,096-token output limit.

Chat

Schematron V2 Turbo

Schematron V2 Turbo is a 3-billion-parameter model from Inference.net built to turn messy HTML into JSON matching a schema you supply, using schema-constrained decoding so the output is valid JSON by construction rather than by retry. It's the throughput-optimized sibling to Schematron V2 Small, handling 4.14 requests per second on a single H100 versus Small's 2.47, a 2.5x gain over the original Schematron 3B. It shares Small's 128,000-token context but doubles its output limit, to 8,192 tokens. On Inference.net's quality benchmark it scores 4.039 out of 5, close to Small's 4.060, and both beat DeepSeek V3.2 and GPT-5.4 Nano on extraction accuracy. On SimpleQA, a web-search factuality benchmark, it reaches 79.42 against Small's 83.10. It suits web scraping and data extraction pipelines that prioritize speed and cost; the endpoint doesn't support tool calling.

Frequently Asked Questions

How do I use ClipTagger 12B?

ClipTagger 12B is no longer available through Puter.js. Explore other AI models for alternatives.

What is the pricing for ClipTagger 12B?

ClipTagger 12B costs $0.3 per 1M input tokens and $0.5 per 1M output tokens.

	Price per 1M tokens
Input	$0.3
Output	$0.5

What is the context window of ClipTagger 12B?

ClipTagger 12B supports a context window of 125K tokens. For reference, that is roughly equivalent to 250 pages of text.

What is the max output length of ClipTagger 12B?

ClipTagger 12B can generate up to 2K tokens in a single response.

Does it work with React / Vue / Vanilla JS / Node / etc.?

Yes — the ClipTagger 12B API works with any JavaScript framework, Node.js, or plain HTML through Puter.js. Just include the library and start building. See the documentation for more details.

Get started with Puter.js

Add AI to your application without worrying about API keys or setup.

Explore Models View Tutorials