Jina AI(@JinaAI_)

Multimodal Search✅ Airgapped✅ Apache2✅

Jina AI (@JinaAI_) — Mon, 08 Jun 2026 07:15:08 GMT

Multimodal Search✅
Airgapped✅
Apache2✅

Han Xiao@hxiao

Built sth I've always wanted, using jina-v5-omni embeddings locally for multimodal file search on mac.
+ airgapped & fully local, indexes text, PDF, image, audio, and video
+ Swift-native UI + mlx-swift-transformer optimized core. No Python.
+ tested on M3 Pro 18G / M3 Ultra 512G / M4 Pro 48G. all work fine, no OOM 
+ local HTTP server exposes index to agents like OpenClaw & Hermes

🔗 View Quoted Tweet

💬0🔄1❤️15👀2718📊3

⚡ Powered by xgo.ing

v5-omni keeps the v5-text backbone completely frozen and adds pretrained vision and audio encoders c...

Jina AI (@JinaAI_) — Tue, 12 May 2026 15:44:42 GMT

v5-omni keeps the v5-text backbone completely frozen and adds pretrained vision and audio encoders connected through small trainable projectors:

- Vision: Qwen3.5 vision encoders with 2x2 spatial merge. We freeze everything except the final projection layer (fc_vision_2), which we replace with a randomly initialized layer mapping into the text backbone's hidden dimension.
- Audio: Qwen2.5-Omni encoder. A single randomly initialized fc_audio layer projects the 1280-dimensional output into the text backbone.
- Video: Handled as a sequence of visual frames, optionally preceded by an extracted audio segment.

💬2🔄1❤️13👀3720📊5

⚡ Powered by xgo.ing

Today v5-omni is available on Elastic Inference Service, HuggingFace and Jina API. Learn more about ...

Jina AI (@JinaAI_) — Tue, 12 May 2026 15:44:42 GMT

Today v5-omni is available on Elastic Inference Service, HuggingFace and Jina API. Learn more about v5-omni from links below.
🤗:huggingface.co/collections/ji…i
arXiv:arxiv.org/abs/2605.083847 
blog:jina.ai/news/jina-embe…W

💬0🔄7❤️26👀3009📊9

⚡ Powered by xgo.ing

jina-embeddings-v5-omni is here! Our first universal embedding model for text, images, audio, and vi...

Jina AI (@JinaAI_) — Tue, 12 May 2026 15:44:41 GMT

jina-embeddings-v5-omni is here! Our first universal embedding model for text, images, audio, and video. Available in two sizes: small (1.57B, 1024-dim, 32K context) and nano (0.95B, 768-dim, 8K context). Both support Matryoshka truncation down to 32 dimensions. 

v5-omni is back-compatible: if you already use jina-embeddings-v5-text-small/nano, the existing text indexes work with v5-omni out of the box. Without reindexing the text, just index your multimodal content with v5-omni and start searching images, audio, and video.

Your browser does not support the video tag.

🔗 View on Twitter

💬11🔄91❤️588👀134532📊204

⚡ Powered by xgo.ing

Our official CLI for agents https://t.co/XLhRvLRuDc

Jina AI (@JinaAI_) — Fri, 13 Mar 2026 12:29:21 GMT

Our official CLI for agents github.com/jina-ai/cli

💬7🔄34❤️223👀21909📊78

⚡ Powered by xgo.ing

Jina AI (@JinaAI_) — Wed, 11 Mar 2026 09:00:14 GMT

Your browser does not support the video tag.

🔗 View on Twitter

💬0🔄1❤️19👀2391📊4

⚡ Powered by xgo.ing

The trend toward smaller embeddings is a shift. On-device retrieval, browser-based search, and edge ...

Jina AI (@JinaAI_) — Thu, 19 Feb 2026 15:24:24 GMT

The trend toward smaller embeddings is a shift. On-device retrieval, browser-based search, and edge deployment all demand models that fit in constrained memory budgets. Learn more about Small & Nano below:
- blog post: jina.ai/news/jina-embe…
- 🤗 weights including GGUFs and MLX:huggingface.co/collections/ji…V
- arXiv:arxiv.org/abs/2602.15547t

💬1🔄2❤️16👀2316📊5

⚡ Powered by xgo.ing

v5-text uses decoder-only backbones with last-token pooling instead of mean pooling. Four lightweigh...

Jina AI (@JinaAI_) — Thu, 19 Feb 2026 15:24:23 GMT

v5-text uses decoder-only backbones with last-token pooling instead of mean pooling. Four lightweight LoRA adapters are injected at each transformer layer, handling retrieval, text-matching, classification, and clustering independently. Users select the appropriate adapter at inference time. For retrieval, queries get a "Query:" prefix and documents get "Document:". Context length is 32K tokens, a 4x increase over v3.

💬1🔄0❤️19👀3307📊4

⚡ Powered by xgo.ing

jina-embeddings-v5-text is here! Our fifth generation of jina embeddings, pushing the quality-effic...

Jina AI (@JinaAI_) — Thu, 19 Feb 2026 15:24:21 GMT

jina-embeddings-v5-text is here! Our fifth generation of  jina embeddings, pushing the quality-efficiency frontier for sub-1B multilingual embeddings. Two versions: small & nano, available today on Elastic Inference Service, vLLM, GGUF and MLX.

💬5🔄16❤️116👀15709📊37

⚡ Powered by xgo.ing

Text embeddings are widely assumed to be safe, irreversible representations. We show we can reconstr...

Jina AI (@JinaAI_) — Wed, 11 Feb 2026 19:35:00 GMT

Text embeddings are widely assumed to be safe, irreversible representations. We show we can reconstruct the original text using conditional masked diffusion.

Existing inversions (Vec2Text, ALGEN, Zero2Text) generate tokens autoregressively and require iterative re-embedding through the target encoder. We take a different approach: embedding inversion as conditional masked diffusion. Starting from a fully masked sequence, a denoising model reveals tokens at all positions in parallel, conditioned on the target embedding via adaptive layer normalization (AdaLN-Zero). Each denoising step refines all positions simultaneously using global context, without ever re-embedding the current hypothesis.

💬1🔄1❤️15👀2437📊5

⚡ Powered by xgo.ing

Check out the live demo https://t.co/W1EXpDFCAL and see it in action. Our read our repo and paper fo...

Jina AI (@JinaAI_) — Wed, 11 Feb 2026 19:35:00 GMT

Check out the live demo embedding-inversion-demo.jina.ai and see it in action. Our read our repo and paper for more technical details on training and decoding.

💬1🔄1❤️14👀2136📊4

⚡ Powered by xgo.ing

Most don't know (1) how easy it is to invert embedding vectors back into sentences, (2) this is a pe...

Jina AI (@JinaAI_) — Wed, 11 Feb 2026 19:34:59 GMT

Most don't know (1) how easy it is to invert embedding vectors back into sentences, (2) this is a perfect task text diffusion models. Here's a 78M parameter model and live demo that recovers 80% of tokens from Qwen3-Embedding and EmbeddingGemma vectors. Works even on multilingual input.

Your browser does not support the video tag.

🔗 View on Twitter

💬7🔄21❤️173👀13024📊65

⚡ Powered by xgo.ing

0.6B params. Top3 on MTEB reranking task. 10× smaller than generative listwise rerankers. Read more ...

Jina AI (@JinaAI_) — Tue, 27 Jan 2026 22:03:03 GMT

0.6B params. Top3 on MTEB reranking task. 10× smaller than generative listwise rerankers. Read more about this Best Paper at AAAI Frontier IR here: arxiv.org/abs/2509.25085

💬0🔄1❤️33👀2357📊6

⚡ Powered by xgo.ing

jina-reranker-v3 was the first listwise reranker to throw all documents into one context window (whe...

Jina AI (@JinaAI_) — Tue, 27 Jan 2026 22:03:02 GMT

jina-reranker-v3 was the first listwise reranker to throw all documents into one context window (where traditional rerankers loop over ⟨q,d⟩ pairs) and let them fight it out via self-attention—what we call "last but not late" interaction. Bold or stupid? But not mediocre. Today it won Best Paper at AAAI Frontier IR Workshop.

💬1🔄14❤️91👀7314📊24

⚡ Powered by xgo.ing

Here's an visualization for jina-embeddings-v4 vectors. This concentration collapses exponents to a ...

Jina AI (@JinaAI_) — Fri, 23 Jan 2026 17:32:03 GMT

Here's an visualization for jina-embeddings-v4 vectors. This concentration collapses exponents to a single dominant value of 127, reducing exponent entropy from 2.6 to 0.03 bits/byte for jina-v4. But bxponent compression alone would yield only 1.1x compression. The additional gain comes from the high-order mantissa byte: when angles cluster around π/2 ≈1.5708, the IEEE
754 mantissa bits encoding the fractional part also become predictable. Empirically, the high-order
mantissa byte entropy drops from 8.0 to 4.5 bits. Together, exponent and mantissa concentration yield the observed 1.5x compression

💬1🔄0❤️25👀4208📊5

⚡ Powered by xgo.ing

Find our C implementation: https://t.co/3fT2TIKs0G And the paper: https://t.co/kdwFRBx18L

Jina AI (@JinaAI_) — Fri, 23 Jan 2026 17:32:03 GMT

Find our C implementation: github.com/jina-ai/jzip-c… 
And the paper: jina.ai/embedding-comp…

💬2🔄4❤️49👀3117📊13

⚡ Powered by xgo.ing

Convert your embeddings to spherical coordinates before compression - this trick cuts embedding stor...

Jina AI (@JinaAI_) — Fri, 23 Jan 2026 17:32:02 GMT

Convert your embeddings to spherical coordinates before compression - this trick cuts embedding storage from 240 GB to 160 GB, and 25% better than the best lossless baseline. Reconstruction is near-lossless as the error stays below float32 machine epsilon - so retrieval quality is preserved perfectly. Works across text, image, and multi-vector embeddings. No training, no codebooks.

💬10🔄65❤️542👀35756📊169

⚡ Powered by xgo.ing

Full paper here: https://t.co/9oWPkIMaW3

Jina AI (@JinaAI_) — Mon, 29 Dec 2025 14:27:21 GMT

Full paper here: jina.ai/pdf/vision-enc…

💬1🔄1❤️19👀1872📊5

⚡ Powered by xgo.ing

Survey on vision encoders in VLMs. The encoder side is weirdly understudied: everyone's busy scaling...

Jina AI (@JinaAI_) — Mon, 29 Dec 2025 14:27:20 GMT

Survey on vision encoders in VLMs. The encoder side is weirdly understudied: everyone's busy scaling the LM while recycling the same 400M CLIP from 2021. We looked at 70+ models and found that training methodology beats scale. A well-trained 400M encoder outperforms a 6B one. Native resolution matters for docs, multi-encoder fusion works. Taxonomy + practical guidance included.

💬2🔄5❤️36👀3620📊12

⚡ Powered by xgo.ing

Jina-VLM achieves state-of-the-art performance among open 2B-scale VLMs, leading with the highest av...

Jina AI (@JinaAI_) — Mon, 08 Dec 2025 07:09:30 GMT

Jina-VLM achieves state-of-the-art performance among open 2B-scale VLMs, leading with the highest average (72.3) across eight general VQA benchmarks, particularly strong on diagrams, charts, and scene text. Its multilingual capabilities stand out most, achieving best-in-class scores on MMMB (78.8) and Multilingual MMBench (74.3).

💬1🔄1❤️11👀2673📊4

⚡ Powered by xgo.ing

Arxiv: https://t.co/s5Sqt1OkOt HuggingFace: https://t.co/HoUcF87BmJ Blog: https://t.co/0xkZlkEte5

Jina AI (@JinaAI_) — Mon, 08 Dec 2025 07:09:30 GMT

Arxiv: arxiv.org/abs/2512.04032
HuggingFace: huggingface.co/jinaai/jina-vlm
Blog: jina.ai/news/jina-vlm-…

💬0🔄1❤️17👀2170📊3

⚡ Powered by xgo.ing

Releasing jina-VLM: our new 2B vision language model achieves SOTA on multilingual visual question a...

Jina AI (@JinaAI_) — Mon, 08 Dec 2025 07:09:29 GMT

Releasing jina-VLM: our new 2B vision language model achieves SOTA on multilingual visual question answering and document understanding among open 2B-scale VLMs.

Your browser does not support the video tag.

🔗 View on Twitter

💬13🔄68❤️648👀53163📊175

⚡ Powered by xgo.ing

curl https://t.co/0iNSUvqigB This is our Meta-Prompt. It allows LLMs to understand our Reader, Embed...

Jina AI (@JinaAI_) — Wed, 30 Oct 2024 15:45:42 GMT

curl docs.jina.ai This is our Meta-Prompt. It allows LLMs to understand our Reader, Embeddings, Reranker, and Classifier APIs for improved codegen. Using the meta-prompt is straightforward. Just copy the prompt into your preferred LLM interface like ChatGPT, Claude, or whatever works for you, add your instructions, and you're set. In this example, we copied the entire prompt into Anthropic Claude and asked it to grab every sentence from Hacker News front page and visualize them using UMAP with matplotlib. This task is nontrivial as it combines multiple APIs from our Search Foundation, like Reader and Embedding where Claude may not have knowledge of. So if you asked Claude directly, it probably wouldn't give an optimal answer. But with the meta-prompt, Claude now has good knowledge about our APIs and can generate much better code! We can copy paste the code directly to Google Colab and with minimum modification, the code just works!

Your browser does not support the video tag.

🔗 View on Twitter

💬15🔄80❤️662👀209063📊106