Superlinked

Stop wasting GPU resources on individual container deployments. Agentic inference requires multiple small models working together, so stop treating them like monoliths. Check out how we built the SIE repo to handle all retrieval modes in one unified process: github.c...

Your inference bill scales with every token you send. At some point, wouldn't it be cheaper to just own the GPU? The answer, from @svonava's Infer() Summit talk: past ~10M requests a month, one GPU running a sub-40B open model beats the per-token API. Self-hosted vs the frontier, on our own numbers: → 50x cheaper (gte-multilingual vs text-embedding-3, all-in on AWS EKS) → 2.7x faster (bge-m3 vs Cohere rerank-3.5) → 96% as smart (Qwen3.6-27B vs GPT-5.1, AA Intelligence Index) → 100% yours, in your cloud or air-gapped The talk walks through the break-even math, why sub-40B open models are now good enough for real work, an $0.80 LoRA that beats a frontier embedder on German legal retrieval, and the two utilization tricks (a shared queue and packing models onto shared GPUs) that keep the GPUs busy. Then it puts it to work: a contract-review agent built on nine small models, all served by one self-hosted cluster. On a real SEC contract it caught a late-delivery penalty written as both 30% and 5%, flagged a missing force-majeure certificate, and marked execution uncertain instead of inventing a signature it couldn't read. Full write-up and the recording: buff.ly/UVO71jV SIE is open source (Apache 2.0): buff.ly/RjhkXiX

Stop overpaying for managed vector embeddings. Self-host on one GPU for 1/10th the price; process 1B tokens every hour. Performance is better too. Grab the SIE repo: github.com/superlinked/sie #AI

Enterprise LLM APIs are a trap. Bills snowball fast at hyperscale. Stop overpaying for inference; grab the SIE repo: github.com/superlinked/sie #AI #LLM

Stop relying on trial and error to fix OOM errors. Most teams just hack batch sizes until it works. The SIE engine changes the game by bounding costs before memory crashes happen. Build better search infrastructure. Try SIE: github.com/superlinked/sie #search #machinelearning #...

Self-hosting models isn't just about saving cash; it is about owning your infrastructure. Pricing changes and token outages disappear when you control the stack. Want to master the math behind inference at scale? Check out the SIE repo: github.com/superlinked/sie #AI #MachineLe...

Stop wasting GPU resources on single model deployments. Smaller, specific models are winning and they perform on par with managed services. Check out the SIE repo to fix your batching and queuing logic. github.com/superlinked/sie #AI #MachineLearning #LLM

Stop fragmenting your AI traffic. Switching to a NATS Jetstream centralized queue helped us boost throughput by 50% on small model workloads. Check out the SIE repo for how we did it: github.com/superlinked/sie #AI #MachineLearning #LLM #Infrastructure

Stop running one container per model. It’s wasting your GPU memory. We built a unified inference engine that lets multiple models live on one GPU using LRU eviction. Check the SIE repo here: github.com/superlinked/sie #AI #LLM #GPU

Stop counting AI requests. It is a trap. Tokens matter more because request size varies wildly. We optimized GPU usage by batching by tokens instead of counts. Get the code here: github.com/superlinked/sie #AI #MachineLearning #LLM

Most contract AI just summarizes text. This agent actually reviews the document. It reads signatures with vision, spots risky inconsistencies, and flags specific obligations. Code is open source here: github.com/superlinked/sie. #AI #ContractTech #OpenSource

Stop running six separate deployments just to build one search pipeline. You are wasting GPUs and sanity. We built SIE to unify everything under one process. Grab the code here: github.com/superlinked/sie #AI #MachineLearning #GPU #SearchEngine #Tech

Stop paying a premium for managed inference. Self-hosting embeddings cuts costs by 90% and slashes latency. Get started with the SIE repo: github.com/superlinked/sie #AI #LLM #Embeddings #MachineLearning

A few months ago, @f_makraduli encountered a bug made his model speak backwards. It turned out to be one of the more instructive things that happened while contributing GPU benchmarks to an open-source research paper. He's talking through the full engineering story at AI Engineer World's Fair tomorrow: weight folding, Triton kernels, what goes wrong when attention becomes bidirectional, and how it all led to a 33-35% speedup on the norm+project operation. "Weight Folding, CUDA Streams, and the Bug That Made My Model Speak Backwards" Catch us at AI Engineer SF, Room 2016, Day 4 · July 2 · 3:45pm ai.engineer/worldsfair/sch… Can't make it, but want to know how you can optimise your inference? Check out our open source github.com/superlinked/sie

We are co-hosting a breakfast with GoodData on June 24th in Prague. Small open source language models are now easy and cheap to self-host, and they match GPT5.1 intelligence. That changes the math on two things most teams are overpaying for: The products you build: Running your own models means owning your ops, infra, and evals instead of renting them. We will swap lessons from large scale deployments. The bill you pay: Plug a self-hosted model straight into your employee agents (Claude Code, Codex, Cursor) and watch the per-token costs drop. If either of those is on your mind and you're in Prague, come and chat with us over coffee. Cut Cost and Build Better Products with Open Source AI Wednesday, June 24, 8:30 to 9:30 AM Rohanské nábřeží 717, Praha 8-Karlín Register here: buff.ly/oDpR48C

We are co-hosting a breakfast with GoodData on June 24th in Prague. Small open source language models are now easy and cheap to self-host, and they match GPT5.1 intelligence. That changes the math on two things most teams are overpaying for: The products you build: Running your own models means owning your ops, infra, and evals instead of renting them. We will swap lessons from large scale deployments. The bill you pay: Plug a self-hosted model straight into your employee agents (Claude Code, Codex, Cursor) and watch the per-token costs drop. If either of those is on your mind and you're in Prague, come and chat with us over coffee. Cut Cost and Build Better Products with Open Source AI Wednesday, June 24, 8:30 to 9:30 AM Rohanské nábřeží 717, Praha 8-Karlín Register here: buff.ly/oDpR48C

Your search stack probably doesn't need more GPUs. You could make efficiency gains with the ones you already have. Modern retrieval runs four model types: dense, sparse, ColBERT, and a cross-encoder reranker. The usual setup gives each its own container and GPU pool, and most of that capacity sits idle while you pay for all of it. In his Berlin Buzzwords 2026 talk, @F_Makraduli shows how to serve all four from one process on a single GPU using SIE, our open-source inference engine. A model like BGE-M3 returns dense, sparse, and multi-vector from one encode call. Reranking runs through the same API. Same server, same GPU. Real BEIR and L4 benchmark data included, plus the honest tradeoffs: when multi-model serving on one GPU wins, and when a model still deserves its own box. If you run hybrid search or RAG, this one is for you! Watch the talk and read the breakdown: buff.ly/FdevCr6

Shopify's product taxonomy goes 8 levels deep with over 10,000 categories. We tested four completely different classification approaches on a 1,790-node slice of it so you don't have to. Our very own Andrey Pikunov has done a systematic evaluation of four approaches to taxonomy classification on Shopify's product hierarchy (1,790 categories, up to 8 levels deep): zero-shot NLI, text retrieval, image retrieval with CLIP/SigLIP, and cross-encoder reranking. Text retrieval with Stella won (hierarchical F1 of 0.425 strict). Reranking didn't help. Image retrieval held up surprisingly well on its own. Full walkthrough: buff.ly/97m4heC