Releasing ColGREP and LateOn-Code models 🚀
ColGREP is a multi-vector search tool built in Rust made for coding agents. It's an hybrid grep which supports both grep features and semantic retrieval. Run 100% locally.
You get two SOTA code retrieval model within ColGREP
Super happy to announce that our paper behind (m)DenseOn & (m)LateOn open recipe is accepted at EACL 2027 🎉🎉🎉 🇬🇷
Cheers to my co-authors and GOATs @raphaelsrty@antoine_chaffin@AmelieTabatta 🐐
A science of post-training can be built on interpretability.
Models know when they reward hack, and we see it, yet training ignores it. We propose Mechanistic Shaping: using the model’s own concepts to write training rules.
In vision, DPO & RL, and its effects on monitoring 🧵
you thought LLMs can only do text optimization and discovery for prompts, harnesses, and code?
turns out they’re starting to be good at natively reasoning about and writing out neural network weights directly too
in the limit, where does this leave symbolic local optimization heuristics like gradient descent…??
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
If you saw Jev and you're wondering how decision models work, this should be helpful! In it I try to explain how Perplexity Decider works, plus some considerations around when and how to use decision models.
Lots of fun excalidraw figures!
Critiquing the classic BEIR regime in IR evals.
Search golden sets are anything but.
They often have sampling biases (ie labeling only BM25 results as good/bad). Until you can measure those, you should take care to limit how much you trust them.
softwaredoug.com/blog/2026/1…
Just one week in, and our new open-source TTS model Phonon has already been downloaded ~85k times! So exciting to bring models to the open-source community. And check out our inference codebase in 🦀, it should make the most out of your local CPU/GPU/NPU/...
The first models of the LFM3 family are out today!! And of course they come with 2 new decision models ⚡
Meet d1-2B and d1-Omni-2B, built on LFM3-VL-2B and LFM3-Omni-2B.
On top of using these much stronger base models, we worked on adding new capabilities that imo make these decision models way more useful. They can now extract text, numbers and dates from a document, draw bounding boxes around objects, count things with a range, and pick several options at once. Still in one forward pass.
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
🐭MICE to Meet You: From Re-ranking Only to End-to-End Retrieval ⚡️
Our #EMNLP2026 paper showed that MICE pushes the Cross-Encoder Pareto frontier.
We take it a step further: turning theoretical pre-computation gains into an operational, end-to-end retrieval architecture 🧵
1/ 💰Multi-vector retrieval shouldn’t charge you by the token!
🎉 Excited to share FLINT (Fixed-budget Late INTeraction), a late-interaction retriever designed for long, complex queries with up to 89% reduction in index and 23-170× faster retrieval latency.⚡
🎛️ FLINT gives a vector-budget knob to balance storage, latency, and quality.
📄 Paper: arxiv.org/abs/2610.27273
How does it work? 🧵
ColBERT stores and compares one vector per token. FLINT pools every query and every document into a fixed number of vectors, so there’s far less to store and far fewer comparisons at search time.
🔎 KURE might be the first major model suite to support all 4 Sentence Transformers model types!
Dense, sparse & multi-vector embeddings, plus rerankers, all for Korean & English.
And damn, the new 1.1B KURE-v3 is really good on English retrieval too 🧵
you can now run perplexity's decider model locally on your mac. it's private, fast, and free, with no api key needed. pplx.ai/decider-for-mac
decider for mac runs pplx-decider-flash-v1.2, our 4B decision model, fully on device. ask a yes/no question about text or an image and get a probability back. after a one-time 5.8 GB download, it serves the decisions api at localhost:8080/v1/decisions.
use it to label files, route messages, check screenshots, or run pre-commit checks, without any data leaving your mac.