GLiClass v4 Edge

Use it to classify text against custom labels, apply decision rules, score responses, or rerank passages on CPU and edge devices.

Availability

Model checkpoint and usage instructions are public on Hugging Face; install with pip or from source.

Pricing model

Not established

Deployment

Built for CPUs and edge devices; examples run locally.

Privacy

Not established

Licence

Apache-2.0

• The 0.4B-parameter model performs instruction-following zero-shot classification. ^[1]
• It returns a score for every supplied label in one forward pass and generates no text. ^[1]

  every label in one forward pass and generates no text.

• The model checkpoint is available as knowledgator/gliclass-v4-edge on Hugging Face. ^[1]

  - The Decision Index results below come from a kit engine (`gliclass_engine.GLiClassEngine`) that applies exactly this mapping, run with `python -m decision_index run --engine gliclass_engine:GLiClassEngine --model knowledgator/gliclass-v4-edge`.

• The release is built for CPUs and edge devices. ^[1]

  **Instruction-following zero-shot classification in a 0.4B encoder, built for CPUs and edge devices.**

• The model supports multi-label and single-label classification, scoring, and reranking. ^[1]
• The model card reports median latency of 11 ms per decision and a Decision Index public skill score of 47.5. ^[1]

  latency of 11 ms per decision, ahead of every model up to 1B parameters on the public board (the best of them scores 36.9).

• DeepSeek opens a real-time web search API Weekly brief 2026-W42
• DeepSeek adds web search for apps and updates its Flash model Daily brief 2026-10-17