Structured output
Uses: /extract · /generate
Two paths to typed JSON: /extract pulls entities and relations straight from text with a small task model, or /generate emits schema-constrained JSON from an LLM. Both return data you can validate.
Featured models
Extract · /extract
| Model | Size | Quality | Latency | Throughput | Cost $/1M | |
|---|---|---|---|---|---|---|
| numind/NuNER_Zero Entities | 449M | 0.6122F1 | — | — | — | |
| urchade/gliner_medium-v2.1 Entities | 195M | 0.6111F1 | 107 ms | 8.9K tok/s | $0.025 | |
| urchade/gliner_small-v2.1 Entities | 60M | 0.5959F1 | 83 ms | 11.7K tok/s | $0.019 | |
| No models match. | ||||||
Measured on L4; other hardware shows "—" until benchmarked. Pick a benchmark to rank by quality.
For similar models, browse the full
/extract catalog →
Generate · /generate
| Model | Size | Quality | Latency | Throughput | Cost $/1M | |
|---|---|---|---|---|---|---|
| Qwen/Qwen3.6-27B MultimodalTool callingConstrained outputStreamingCodeSQL | 27.0B | 0.6000acc | 1.7 s | 222 tok/s | $3.80 | |
| Qwen/Qwen3-4B-Instruct-2507 Long contextTool callingConstrained outputStreamingCodeSQL | 4.0B | 0.6033acc | 576 ms | 472 tok/s | $1.78 | |
| Qwen/Qwen3-0.6B Streaming | 600M | 0.4600acc | 413 ms | 595 tok/s | $1.41 | |
| No models match. | ||||||
Measured on RTX-PRO-6000; other hardware shows "—" until benchmarked. Pick a benchmark to rank by quality.
For similar models, browse the full
/generate catalog →
Examples
End-to-end projects from our examples that put this task to work.
Swap an OCR model with one identifier change
A recognition VLM, an end-to-end document model and zero-shot NER, all driven by the same extract call.
Self-hosted product search in 5 min Amazon-style search from three SDK calls: extract attributes, encode descriptions, score-rerank candidates.
Multimodal product classifier with embeddings NLI, text and image retrieval, and cross-encoder reranking over a hierarchical product taxonomy.
Stripe checkout with a fraud-risk gate extract + encode + score every order against a corpus of past fraud patterns before the payment is created.
Featured picks are still being finalized. Latency, throughput and cost are real where we've benchmarked the model on the selected GPU; "—" means no measurement there. Cost is approximate — computed from list GPU prices; your actual price depends on the provider you deploy SIE with.
Compare (0)Compare →