Interfaze
- title
- Interfaze
- type
- entity
- summary
- AI startup building a deterministic-by-design model aimed at structured output and parsing tasks; published the open Structured Output Benchmark in April 2026
- tags
- company, llm, structured-output, benchmark
- created
- 2026-04-30
- updated
- 2026-04-30
Interfaze (interfaze.ai, currently in beta) is building a model focused on deterministic structured-output tasks — parsing invoices, medical records, transcripts, OCR, PDF-to-database. Their pitch is that general-purpose frontier models do this work badly enough that a specialized model trained against the right metrics is worth a separate product.
In April 2026 they published the Structured Output Benchmark (SOB) — see structured-output-benchmark — with arXiv 2604.25359, a HuggingFace dataset, and a hosted leaderboard. The benchmark places their own Interfaze-Beta model at #6 (Overall 0.855) on the unified leaderboard, behind GPT-5.4, GLM-4.7, Qwen3.5-35B, Gemini-2.5-Flash, Qwen3-235B, and ahead of Claude-Sonnet-4.6, GPT-5, Gemma-3-27B, and others.
What's interesting beyond the model
The benchmark is the more durable contribution. By splitting the score into seven metrics across three modalities (text/image/audio) and weighting by schema complexity, SOB exposes the json-pass-value-accuracy-gap — JSON Pass Rate clears 95%+ on every frontier model, but exact leaf-value accuracy sits 15-30 points lower. That gap is the single most useful corrective to "structured output is solved" claims circulating in 2026.
Interfaze's framing of their own positioning is honest about the conflict: they're publishing a benchmark on which their product will be measured. The implicit bet is that being credibly measured is worth more than gaming a private internal score.
Coverage in the wiki
- structured-output-benchmark — full coverage of SOB, the methodology, and the headline numbers
- json-pass-value-accuracy-gap — the standalone concept that's the benchmark's load-bearing finding