Interfaze 1 Lite
An Apache-2.0 open-weight model from the Hugging Face org interfaze-ai, billed as a "mixture-of-architectures" vision-language reasoning model for developer workloads. The model card is the only source here (T3: first-party card, no independent eval, no known org details).
What it is
It is not one network. A Qwen-based reasoning core calls specialist sub-models through tool calls, and the card names them:
- Chandra OCR for documents, PDFs and Word files, plus PaddleOCR
- Whisper with pyannote for transcription and speaker diarization
- SAM 2.1 for visual grounding and segmentation
- TimesFM for time-series forecasting
- Llama Guard for safety moderation
So the weights bundle other labs' models behind one interface. That is the same move sakana-fugu makes with hosted frontier models, except here every part is open and it runs locally.
Claimed capabilities
- OCR on images and PDFs (up to 50 pages), with tables and layout boxes and per-line confidence.
- Timestamped transcription with diarization; the card says about 90 seconds for a 95-minute recording.
- Open-vocabulary object detection, GUI element location for agents, JSON-schema-constrained output.
- Translation across 160+ languages; reasoning across science, math, SQL and 14+ languages, with 131k context.
Reported scores: 85.9 on GPQA Diamond and 83.8 on olmOCR-Bench. Both are self-reported. The card gives no comparison table here, so neither number can be ranked against qwen or other open models on this evidence.
Access and limits
Runs on one 80GB GPU, offline. Locally through Hugging Face Transformers with trust_remote_code=True, or through a hosted OpenAI-compatible Interfaze API (no pricing recorded). The card's own limitations: slow generation, no response schema in chat, and heavy memory use on large PDFs.
Why it matters here
It sits on the open-weight-models map as a composition product rather than a new base model. The "architecture has stopped corresponding to a lab" thread in synthesis gains another point: Qwen reasoning core plus Chandra, Whisper, SAM and others, under a new name.
Provider-level context: llm-provider, llm-benchmarks.