On May 11, 2026, the team behind Interfaze (whose benchmark repo lives under the JigsawStack organization) published a new model architecture under the same name, and the Hacker News thread picked up 164 points and 43 comments. The core claim is blunt: when a generalist LLM does OCR, speech-to-text, or web extraction, it is spending a many-billion-parameter reasoning engine on a job a small specialized network could handle. Interfaze stitches the two together — specialized DNN/CNN components for the task-specific work, an omni-transformer for decoding and general understanding, plus task adapters and built-in infrastructure: a web index, scraping, and a code sandbox.
This is not another “matches the frontier” story. It is a direct attack on the flash/mini pricing tier, aimed at teams that push millions of tokens a month through extraction pipelines on cheap general models.
A Hybrid: Specialized Nets Plus an Omni-Transformer
Interfaze puts task-specialized DNN/CNN components and a transformer decoder inside one model, with adapters and infrastructure (web index, web scraping, code sandbox) as native capabilities rather than bolted-on tools. The specs: a 1M-token context window, 32k max output, inputs across text, images, audio, and files, and a reasoning mode that is optional and off by default — for deterministic tasks, reasoning mostly adds latency and cost. Pricing is $1.50 per million input tokens and $3.50 per million output tokens, the same band as Gemini-3-Flash. The API endpoint is OpenAI Chat Completions-compatible, the model currently runs as interfaze-beta, and there are TypeScript and Python SDKs plus Vercel AI SDK and LangChain integrations.
Nine Benchmarks, All Self-Reported
The announcement ships nine benchmarks against Gemini-3-Flash, Claude-Sonnet-4.6, GPT-5.4-Mini, and Grok-4.3. The numbers are eye-catching. OCRBench V2: 70.7% against 55.8% for Gemini-3-Flash. olmOCR: 85.7%, ahead of Grok-4.3 at 81.9% and also above Chandra OCR 2 (84.3%), Reducto (76.2%), and Mistral OCR (72.0%). RefCOCO object grounding: 82.1%, with Claude-Sonnet-4.6 at 75.5% and Grok-4.3 collapsing to 25.0%.
On speech, VoxPopuli word error rate lands at 2.4%, second only to Scribe v2 at 1.7%. On text and reasoning: Spider 2.0-Lite at 52.9% versus Claude’s 49.6%; GPQA Diamond at 89.9%, tied with Claude-Sonnet-4.6; MMMLU across 14 languages at 90.9% versus Grok-4.3’s 89.7%; MMMU-Pro at 71.1% versus 68.7%. The final entry, SOB (Structured Output Benchmark), reads 79.5% against 78.4% — and SOB is Interfaze’s own benchmark, released the week before. To be clear about the caveats: all nine are self-reported (the benchmark repo is on GitHub), and the comparison set is deliberately flash/mini-tier, not frontier models.
Partial Activation and Precontext
Two design choices stand out. The first is partial activation: tag the system prompt with a task such as ocr, and only the relevant weights run, producing deterministic output at lower cost. The second is Precontext: alongside the schema-formatted answer, the model returns raw per-line and per-word bounding boxes with confidence scores. For anyone verifying OCR results or running quality gates inside a document-extraction pipeline, that metadata is far more useful than a bare text answer.
The speed numbers are also strong: 209 seconds of audio transcribed per second of compute, meaning a 1h35m podcast finishes in about 50 seconds — roughly 1.5x faster than Deepgram Nova-3, about 8x faster than Scribe v2, and 11x or more faster than Gemini-3-Flash.
Should You Route Workloads to It
The decision framework is simple. If your product already runs serious volumes of OCR, transcription, or structured extraction on flash-tier models, Interfaze’s pricing and speed justify a bake-off on your own data — trust your own evals before any leaderboard. But read the risk list too: the weights are closed, the model is in beta, every benchmark is self-reported, and the headline structured-output benchmark is the company’s own. Across the 43 Hacker News comments, skepticism about the “hybrid architecture” marketing language and the self-reported numbers was common. That skepticism is a healthy signal, not a verdict.
In the bigger 2026 picture, this is another data point in the split between specialized and generalist routes (a divergence we flagged in our opening-of-the-year outlook): as the marginal cost of generalist models fails to fall fast enough, carving deterministic tasks off to task-specialized architectures becomes an ordinary engineering decision rather than an exotic one.
Sources
- Interfaze: A new model architecture built for high accuracy at scale — Interfaze
- interfaze-complete-benchmarks — GitHub
- Interfaze: A new model architecture built for high accuracy at scale — Hacker News
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.