🤗 Ling-3.0-flash is now also open weighted! - MoE with 124B total & 5.1B active params - Hybrid linear attention for higher efficiency with longer sequences Overall very competitive for its size! Nice work to the Ant Group who worked on this.

We now have an MTEB account that posts about new embedding & related models! Give them a follow to stay up-to-date on the latest releases.

MTEB@mteb.org · 2d ago

New on MTEB: hotchpotch/bekko-embedding-v1-a8m is mmBERT-small with its 22 layers pruned to 4 — active parameters cut from 42M to 7.7M, vocabulary left intact. It scores 56.73 on MTEB(Multilingual, v2), against 47.21 for a static model of the same total size.

A film catalogue tells you what a film is about, not what happens inside it. So: 1,864 public-domain Prelinger films, described every ~60 seconds by an open 2B video model. 23,148 searchable moments. Search "typing on a computer keyboard", land on the second it happens.

🔧 Sentence Transformers v5.6.1 is out: a patch release fixing silently degraded flash attention embeddings for XLM-R & RoBERTa models. If you encode with flash_attention_2 on transformers v5, you'll want this one 🧵

🎉 @lightonai.bsky.social just published LightOn-rerank: rerankers that score text passages or document page images against a query. Six models: Qwen3.5 at 0.8B / 2B / 4B, each in a pointwise and a generative listwise variant. Excellent for text <-> image retrieval. 🧵

I'm very excited to share that NVIDIA just released Nemotron-3-Embed: two multilingual embedding models for retrieval The 8B takes the top spot on RTEB. - Nemotron-3-Embed-1B-BF16 (2048-dim) - Nemotron-3-Embed-8B-BF16 (4096-dim) Both OpenMDW-1.1, ready for commercial use. 🧵

Tencent just published R3-Skill, a two-stage retrieval stack purpose-built for a problem RAG-style retrievers weren't designed for: routing LLM agent skills (think Anthropic's SKILLmd format). Two 0.6B models, both Apache 2.0, one embedding model, and one reranker. 🧵

NAVER, the original authors of SPLADE, just published V-SPLADE, an inference-free sparse retriever for visual document retrieval. Two models on the same backbone (ModernVBERT, 250M params), both Apache 2.0: - naver/v-splade-quality - naver/v-splade-efficient 🧵

💧 Liquid AI released 2 multilingual retrieval models, the first bidirectional members of the LFM family. Both 350M params, 11 languages (ar, de, en, es, fr, it, ja, ko, no, pt, sv): - LFM2.5-Embedding-350M (bi-encoder) - LFM2.5-ColBERT-350M (multi-vector, late interaction) 🧵

🐛I've just released Sentence Transformers v5.6.0! A correctness- & robustness-focused release, headlined by a fix for a silent scoring bug in causal-LM rerankers (think Qwen3-Reranker), plus a batch of hard-negative mining & loss-correctness fixes. Thread with highlights 🧵

The core Massive Text Embedding Benchmark (MTEB) team has heavily updated their leaderboard Space. It's so, so much faster, and there's much more information about the best embedding models (and related) to glean from it. Details in 🧵

LAION just released VoiceCLAP-Large-v2, a contrastive voice-text embedding model that's essentially CLIP for voice and emotion. 9B params via a rank-16 LoRA on top of LCO-Embedding-Omni-7B (itself based on Qwen2.5-Omni thinker). Apache 2.0. 🧵

🤗 Announcing the Ettin Reranker family: six new CrossEncoder rerankers from 17M to 1B parameters, state-of-the-art at their respective sizes. Built on the Ettin ModernBERT encoders, with the full training recipe and ~143M-triple training dataset as well. 🧵

Fastino Labs just released GLiGuard, an open-source safety moderation model that remembers encoders are king for these kinds of tasks. One model, Apache 2.0: gliguard-LLMGuardrails-300M: 300M params, evaluates multiple safety tasks at a time. 🧵

🤖 I've just released Sentence Transformers v5.5.0! It's headlined by a new `train-sentence-transformers` Agent Skill: let your AI coding agent train & finetune embedding, reranker & sparse encoder models for you. Plus more, here's a thread with highlights 🧵

The excellent zerank-2 reranker model by @zeroentropyai.bsky.social is now fully compatible with Sentence Transformers, no `trust_remote_code=True` needed. It's 4B and cc-by-nc-4.0, and performs very well. I'm quite fond of their training methodology, I'll explain in the 🧵

IBM just released the R2 generation of their Granite multilingual embedding models for retrieval, and the jump over R1 is very notable. Two models, both Apache 2.0: - granite-embedding-97m-multilingual-r2 (384-dim) - granite-embedding-311m-multilingual-r2 (768-dim) 🧵

Dun Zhang just released Prism-Reranker: a new open-source reranker family for improving (agentic) search on Qwen3.5, in four sizes from 0.8B to 9B. MIT licensed. Drop-in CrossEncoder for scores only. Trained on a realistic agent-style data mix. 🧵

BidirLM-Omni-2.5B-Embedding is live: a single bidirectional encoder that embeds text, images, and audio into the same space! Three modalities, all in one 2048-dim space. 🧵

The @flowercomputer.com team brought Static Embedding models to Rust and carried out optimizations to heavily speed up potential inference. They'll be open sourcing the deployment pipeline soon. This looks very nice at first glance! Definitely give the thread a read.

é.@urcad.es · 4mo ago

We've recently implemented the fastest static embedding model in the world by an insanely large margin Read on for additional info ↓

📈 New blog post: Training and Finetuning Multimodal Embedding & Reranker Models with Sentence Transformers. As a practical example, I finetuned Qwen3-VL-Embedding-2B for Visual Document Retrieval (matching text queries to document screenshots). Thread with highlights 🧵

This is what finetuning is all about: hyper-specific use cases tackled using a custom finetuned model to blow the generic baselines out of the water! Nice work to Aresh Tajvar on this San Diego Municipal Code retrieval model.

🔧 Sentence Transformers v5.4.1 is out, a small patch release patching support for numpy string arrays & improving the safety of activation function loading. Two fixes below 🧵

🧩 To celebrate yesterday's Sentence Transformers v5.4 release, I went back to update SpanMarker: my Named Entity Recognition project. It's still a solid, extremely efficient option for NER. Here's how it works and what's new 🧵

🌐 I've just released Sentence Transformers v5.4: we're going fully multimodal for embeddings & reranking! Also featuring a modular CrossEncoder, and automatic Flash Attention 2 input flattening. Highlights in 🧵

⬆️ I've just released Sentence Transformers v5.3.0! This release upgrades training with MultipleNegativesRankingLoss with alternative InfoNCE formulations and hardness weighting, adds two new losses, and more. Details in 🧵

🤗 Perplexity has released 4 open-weights state-of-the-art multilingual embedding models designed for retrieval tasks! pplx-embed-v1 and pplx-embed-context-v1 Specifically trained for int8 and binary embeddings, they'll be viable for massive search problems. Details in 🧵

🚀 LightOn is back with a SOTA late-interaction model for search: ColBERT-Zero! By performing contrastive pre-training directly in the multi-vector setting, it outperforms GTE-ModernColBERT etc. on BEIR, using only public data and reaching 55.43 nDCG@10. Details in 🧵

ggml / llama.cpp are joining @hf.co, ensuring it'll stay open, maintained, and up to date for a long long time! 🚀 huggingface.co/blog/ggml-jo...

👏 Jina AI is back with new state-of-the-art multilingual embedding models for retrieval & more: jina-embedding-v5-text! 2 efficient sizes, 239M & 677M, they outperform Qwen3-embedding, EmbeddingGemma-300m, multilingual-e5-large, etc. Details in 🧵