Marcin Antas

Weaviate Core Engineer

[Marcin Antas @antas_marcin]

@GPW_Trader2022 250

[Bob van Luijt @bobvanluijt]

🔥 The Weaviate team is on absolute fire this month, shipping one major update after another!

🗣️ Today, I am incredibly excited to roll out a feature requested by MANY of our community members

🚀 We have just launched a 𝗙𝗥𝗘𝗘 𝗙𝗢𝗥𝗘𝗩𝗘𝗥 𝗧𝗜𝗘𝗥 on Weaviate Cloud!

[Marcin Antas @antas_marcin]

@helloiamleonie @liquidai @maximelabonne @paulabartabajo_ wow! Congratulations!!!

[Connor Shorten @CShorten30]

You can use AgentIR embeddings in the Weaviate Database with the `text2vec_huggingface` module!
🤗💚

And Happy Birthday to the lead creator and maintainer of Weaviate Modules, @antas_marcin! 🎂

[Marcin Antas @antas_marcin]

@CShorten30 Awesome!! and thank you for birthday wishes @CShorten30 !!!

[Marcin Antas @antas_marcin]

@KryptoDzikPL Super seria postów. Bardzo ciekawa. Mam pytanie co sądzi Pan o nadchodzącym IPO Cerebras?

[Marcin Antas @antas_marcin]

@CShorten30 @MarcusForPeace @GoogleResearch NIce!!!

[Marcin Antas @antas_marcin]

@KinasRemek @nvidia Mega! Gratulacje!!!

[Marcin Antas @antas_marcin]

Weaviate on NVIDIA GTC 2026 slides!

[Bob van Luijt @bobvanluijt]

👀 And what did my eyes see during Jensen's #GTC keynote...?

[Weaviate AI Database @weaviate_io]

The era of juggling 5 different embedding models is over.

Google just unified text, images, video, audio, and PDFs into one vector space.

𝗢𝗻𝗲 𝗺𝗼𝗱��𝗹, 𝗺𝘂𝗹𝘁𝗶𝗽𝗹𝗲 𝗺𝗼𝗱𝗮𝗹𝗶𝘁𝗶𝗲𝘀: Text, images, video, audio, and PDFs all mapped into a single unified vector space. No more juggling different embedding models or complex preprocessing pipelines.

𝗕𝘂𝗶𝗹𝘁 𝗼𝗻 𝗚𝗲𝗺𝗶𝗻𝗶 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 with support for 100+ languages and some impressive specs:
• 8192 max input tokens
• Flexible output dimensions (128-3072)
• Top 5 performance on MTEB Multilingual leaderboard
• SOTA among proprietary models across most modalities

𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗳𝗼𝗿 𝘆𝗼𝘂𝗿 𝗥𝗔𝗚 𝗮𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀:
By natively handling interleaved data without intermediate processing steps, Gemini Embedding 2 simplifies complex pipelines. You can now build semantic search and recommendation systems that seamlessly work across text documents, images, videos, and audio files.

The model is available now via Gemini API and Vertex AI, and works with Weaviate's existing text2vec-google integration 💚

Check out these recipes to get started 👇

Semantic search/RAG over video: https://t.co/IzU7XZ34N7

Semantic search/RAG over audio: https://t.co/q7WNOncgx4

Multimodal PDF RAG: https://t.co/hPhbcNwk4D

[Etienne Dilocker @etiennedi]

Small Change, Big Impact (Day 3/5): 12x Reduction of Inter-zonal traffic ⚡️📉

Another day, another significant win. Today it's not a performance update, but a traffic optimization.

Learn how we reduced traffic by 12x on a large customer's cluster with 1,700+ updates per second. 🧵

[Etienne Dilocker @etiennedi]

Small Change, Big Impact (Day 2/5): Faster re-scoring for compressed HNSW (PQ/SQ/RQ/BQ) ⚡️📈

Any time you have compression, you have rescoring. Today's hidden improvement is a 25%+ speed-up from faster rescoring through better re-use of resources. Details in 🧵

[Etienne Dilocker @etiennedi]

Small Change, Big Impact: Day 1/5: PQ Speed-up 📈⚙️

Product Quantization just got ~60% faster on average between v1.34.7 and v1.34.8.

How? Why? It uses an optmization technique that's probably as old as coding itself. More in 🧵⬇️

[Bob van Luijt @bobvanluijt]

💪 Nice! Awesome work, Weaviate team!
👉 https://t.co/gYlVVxdXjT

[Marcin Antas @antas_marcin]

If you want to have good multi lingual text and image search then you should use OpenCLIP models, for only image search I would go with ModernVBERT-embed.

I have used this dataset: https://t.co/DzrGXovC2p

and made the whole project open source, so if you want to run it locally the checkout this repo: https://t.co/utrWvYdzFe

[Marcin Antas @antas_marcin]

Want to compare different CLIP models head-to-head?

I've built a simple web app where you can test 3 search types (similarity search, image search, text-to-image search) across 4 open-source CLIP models:

- facebook/metaclip-2-worldwide-b32-384
- ModernVBERT/modernvbert-embed
- OpenCLIP xlm-roberta-base-ViT-B-32 pretrained: laion5b_s13b_b90k
- google/siglip2-so400m-patch16-512

Images are stored in @weaviate_io and inference runs on an NVIDIA Jetson AGX Orin.

Try it out and evaluate them yourself: https://t.co/lnSVP1Hx8I

#weaviate #nvidia #agxorin #clip

[Marcin Antas @antas_marcin]

@KinasRemek Zapowiada się mega ciekawie!

[Marcin Antas @antas_marcin]

@RBrzoska @MistralAI @bielikllm to by było coś niesamowitego!

[Marcin Antas @antas_marcin]

With our newest CLIP inference container we added support for NVIDIA Jetson devices (JetPack 6) here's link how to use it: https://t.co/cUB0lB4z9m

[Weaviate AI Database @weaviate_io]

Just released: Multi2Vec CLIP inference container 1.5.0 🎉

This release contains:
- Support for facebook MetaClip2 models
- Support for ModernVBERT/modernvbert-embed model
- Added support for running inference container on NVIDIA Jetson devices

Check out the docs to spin it up: https://t.co/08dr4zTaIt

[Weaviate AI Database @weaviate_io]

Just released: Multi2Vec CLIP inference container 1.5.0 🎉

This release contains:
- Support for facebook MetaClip2 models
- Support for ModernVBERT/modernvbert-embed model
- Added support for running inference container on NVIDIA Jetson devices

Check out the docs to spin it up: https://t.co/08dr4zTaIt