Clement Delangue

Articles & links

RT @ZixuanLi_: GLM-5.2 is now @Zai_org's most-liked model on Hugging Face of all time. https://t.co/qFgxVrzer8 https://t.co/SIS5y0BEN1

Models – Hugging Face huggingface.co
AI Weekly's analysis
  • Stability AI's stable-video-diffusion-img2vid-xt leads the most-liked board at 3.33k likes, with OpenAI's GPT-2 a step behind at 3.32k.
  • DeepSeek-OCR, dated November 4, 2025 on the page, has already reached 3.29k likes and 2.21M downloads.
  • Google's BERT-base-uncased posts 61.2M downloads on the page, with BAAI's bge-m3 next visible at 31.1M.
Read full analysis →
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 2 from the directory shared this · 20d ago

Rampart from the @ndstudio @WhiteHouse is number one trending token classification model on HF. Very cool to see public organizations starting to own and build their weights instead of renting them from an API provider! https://t.co/XhszxoMOx8 https://t.co/pule1rVvsa

Token Classification Models – Hugging Face huggingface.co
AI Weekly's analysis
  • A 1B model named privacy-filter, published under the openai account, shows 288k downloads and 1.68k likes on the Hugging Face token classification trending list.
  • fastino's gliner2-privacy-filter-PII-multi, a 0.3B multilingual PII detector, sits high in trending with 41.9k downloads about ten days after its update.
  • Hugging Face lists 28,597 token classification models in total, with dslim's bert-base-NER still leading by lifetime pulls at 1.88M downloads.
Read full analysis →
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 18d ago

Super excited about open-source router systems and routing models like @vllm_project semantic router: https://t.co/YtSr19nSaS The future is multi-models and you'll want to customize your router the same way you customize your code! It could be the key to tilt the value capture…

llm-semantic-router (vLLM Semantic Router) huggingface.co
AI Weekly's analysis
  • The vLLM Project released Semantic Router v0.1 'Iris' on January 5, 2026, with contributions from over 50 engineers across Red Hat, IBM Research, AMD, and Hugging Face.
  • The router extracts six signal types (domain, keyword, embedding, factual, feedback, preference) and composes them through a configurable decision engine, replacing an earlier 14-category approach.
  • Since its September 2025 launch the project reports over 600 merged pull requests and 300+ closed issues, with the Hugging Face org now hosting 58 models and 13 datasets.
Read full analysis →
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 20d ago

RT @ndstudio: HuggingFace: https://t.co/0P64HHeVDa

nationaldesignstudio/rampart · Hugging Face huggingface.co
AI Weekly's analysis
  • Rampart is a 14.7 MB ONNX token-classification model that redacts PII in user-typed text before it leaves the browser.
  • On a 30,000-row test set across seven Latin-script languages, the card reports 98.42% private-term recall and 91.69% public retention.
  • Reported p50 latency is 3.9 ms on WebGPU and 12.6 ms on WASM, but non-Latin script recall is around 13.7%.
Read full analysis →
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 21d ago

Kog open-sourced on @huggingface the 2B model that they used to show a model running at 3,000+ tokens per second. Very cool work! https://t.co/fjCnAwQoWe https://t.co/k8hD7xW0F7

Kog Laneformer 2B: The Latency-First Model Behind Kog Inference Engine huggingface.co
AI Weekly's analysis
  • Kog released Laneformer 2B, a 2.3B-parameter instruction-tuned coding model built around decoding speed rather than benchmark score.
  • The team reports 3,000 output tokens/s on 8× AMD MI300X and 2,100 on 8× NVIDIA H200 at FP16, batch size 1.
  • Laneformer 2B scores 45.1% on HumanEval+ and 51.6% on MBPP+ in greedy decoding, with sliding-window attention on 10 of 15 layers.
Read full analysis →
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 26d ago