This architecture unlocks new use cases we couldn’t reach before, and it’s just the beginning. We’ll keep pushing the frontier of joint encoder-decoder information extraction. Let us know what you build with it! 🚀. #NER #NLP #OpenSource #LLMs #GLiNER.

Check our blog post for more technical details:. 📖

The most fun image & video creation tool in the world is here. Try it for free in the Grok App.

We’re releasing 3 model sizes powered by:.🔹 DeBERTa v3 (encoder).🔹 SmolLM2-135M (decoder).And yes — it’s all open source 👐.

What new opportunities does it open?. 💡 No need for fixed label sets — works with open ontologies. 🧠 Plug-and-play decoder expands knowledge easily. 🧩 Supports more than NER: now includes entity linking, descriptions, and beyond.

Unlike typical pipeline approaches, our new GLiNER-decoder architecture does everything once and all in latent space. It fuses the contextual strength of encoders with the knowledge-rich power of decoders — without compromising on speed.⚡

🚀 GLiNER x SmolLM: a new joint encoder-decoder architecture 🚀. We are excited to release a new kind of GLiNER model built with the mantra "you do the same things only once.". Built on top of DeBERTa + @huggingface SmolLM2 — full details below 👇.

Special thanks to @ihor_step , @bioMikeee , Dmytro Vodianytskyi, and Oleksandr Lukashov for their outstanding contributions, as well as @urchadeDS for the inspiration and encouragement that shaped this research.

Check the GitHub repo here: .📦 GitHub:

The models are fully open-source:. 🤗 Hugging Face:

🧩 Post-training with LoRA adapters improves specialization and stability in smaller models, with higher-rank configurations proving particularly effective for the edge variant.

🎯 Few-shot adaptation yields up to +50% performance gain with only 8 training examples per label

🤖 We adapted Reinforcement Learning (PPO) to multi-label text classification — refining decision boundaries and improving label–text alignment

⚡Developed models maintains 80% throughput even with 128 labels (vs. ~52× slowdown for cross-encoders)

📊 Out models outperforms strong cross-encoder baselines by +5.5%, while running 2.3×–16× faster

The earliest GLiClass models performed well with small label sets but struggled to generalize in broader scenarios. We systematically addressed these limitations, introducing architectural refinements, reinforcement learning–based training, and targeted post-training methods,.

🚀 Our largest study on zero-shot text classification is out!.📄 We surpass cross-encoders while being much faster, especially for large label sets. Check out all the research results 👇.

RT @ClementDelangue: When you realize that open-source is at the frontier of AI despite:.- less GPUs.- less money.- less public and policy….

📦 Find the GitHub repo here.👉 Massive thanks to our amazing team! 💪.

🧠 Variants include:.The models are based on DeBERTa, ModernBERT and the Ettin small model for edge device use-cases. – gliclass-edge-v3.0: ultra-efficient.– gliclass-large-v3.0: high accuracy.– gliclass-x-base: robust multilingual zero-shot.

⚡ Efficiency.The models are much faster compared to traditional cross-encoders, especially in the case of a large number of labels. – Up to 97 examples/sec on A6000 (batch=1).– Handles long context and 100+ labels.– Edge-ready speeds with strong generalisation