←
Superlinked976 posts
Superlinked avatar✓

Superlinked ✓

@superlinked

SIE serves small open models for everyday AI tasks. Find a model that meets your quality target, then call our API or run it in your own cloud.

◷ Joined September 2019▧ 397 Photos and videos
Superlinked avatar
Superlinked@superlinked·Nov 3

Help cut your AI costs with models built for the task! Run 100+ task-specific models behind one production API. Check out SIE Serverless in @svonava 's post below.

1:05
Superlinked avatar
Superlinked@superlinked·1h

Getting an enterprise AI project into production isn’t just about choosing the best model or having enough budget. As we head into 2027, there’s another question companies are taking much more seriously: Where does our data actually go? Think about it. A single AI agent might call several different models to complete one task. Along the way, sensitive information like customer records, contracts, or even source code could end up passing through multiple third-party providers. And that’s where things get tricky. So, how do you keep that data safe without giving up the benefits of AI agents? Check out the deep-dive here: medium.com/towards-data-engi…

Post attachment
Daniel Svonava avatar
⟳ Superlinked retweeted
Daniel Svonava@svonava·20h

Let's fix the marginal utility of your next token 💵 There are still folks that don't saturate all their coding subscriptions. In my experience, this also correlates with being unhappy with agentic development in general or not having the right mind-set about it yet 👇

6:54
Superlinked avatar
Superlinked@superlinked·Nov 17

Take back control of your AI stack with small-model inference! Join @svonava on Nov 19th for a webinar with @gooddata_ai on how you can bring inference in-house and build a truly sovereign AI product that doesn't send data via third party APIs. Sing up here: luma.com/n7l3apll

0:58
Vipin Gautam (Viipin I Gautam) avatar
⟳ Superlinked retweeted
Vipin Gautam (Viipin I Gautam)@viipin8·Nov 13

Your AI agent shouldn't waste expensive tokens reading PDFs. Most agents send everything to frontier models, even tasks that open models can handle just as well. That's where Superlinked Inference Engine (SIE) comes in. It lets Claude and other MCP clients offload these tasks to your own cluster: → PDFs & Documents: Turn PDFs, Office files, and scans into clean Markdown. → Images: Handle captioning, object detection, and visual Q&A. → Structured Data: Extract schema-valid JSON without wasting frontier-model tokens. Powered by open models like LightOnOCR, MinerU, Docling, Florence-2, GLiNER2, and Qwen. One MCP edge. Multiple inference workloads. Your own infrastructure. Open source under Apache 2.0. GitHub: github.com/superlinked/sie If you're building with Claude or MCP clients, this deserves a star. ⭐

Post attachment
Superlinked avatar
Superlinked@superlinked·Nov 16

Revolut switched to open models this year and saved 10x. They still use frontier models, but most of the task-specific work is now done by GLM and other open source models. In our opinion, this is the future of how all enterprises will run their AI stack. Reliance on 100% frontier for all work is simply too unreliable and wasteful to last as a long-term AI strategy. AI and data sovereignty is going to become a much larger topic as people look to fully control their models and drive down token bills. You can run 100+ open models behind one API with SIE Serverless, or completely self-host with our OSS: Serverless: superlinked.com/serverless Github: github.com/superlinked/sie Read the story on Revolut's open models adoption here: theregister.com/ai-and-ml/20…

Post attachment
Daniel Svonava avatar
⟳ Superlinked retweeted
Daniel Svonava@svonava·Nov 15

Is your agent chained to a specific environment? What if a client wants to pay you the big bucks to run the whole thing in their cloud - including the models? We will discuss with @gooddata_ai AI CTO how they deal with this on Nov 18th, online. Comment below / DM for an invite.

0:51
Daniel Svonava avatar
⟳ Superlinked retweeted
Daniel Svonava@svonava·Nov 16

RLMs aren't just a long-context trick. They're a KV-cache trick. Same Qwen3.8-27B on one RTX PRO 6000: ~4 concurrent requests with the full 262K trace in context, ~32 with an RLM over 32K slices because of their prefil / decode biases:

5:08
Superlinked avatar
Superlinked@superlinked·Nov 13

Sovereign AI starts with inference. But what does bringing inference in-house actually look like? More enterprises want control over their AI stack. That means keeping sensitive data secure, reducing dependence on external model APIs, and making inference costs more predictable. We're teaming up with GoodData and Demetrios Brinkmann from the Agentic AI Foundation to explore how smaller, specialised models can make this possible. We'll get into: 🔹 GoodData's journey: Why they're bringing inference in-house, the challenges they've faced, and what they've learned about model evaluation and deployment. 🔹 Small model composition: Using classifiers, guardrails and confidence-based routing to reduce reliance on larger LLMs. 🔹 Live technical demo: Daniel will walk through the architecture, code and serving stack behind running multiple models with Superlinked's Inference Engine (SIE). No hypothetical architectures or AI hype. Just an honest look at the engineering decisions, trade-offs and infrastructure involved. Speakers: Zdeněk Musil, CTO, AI at GoodData Daniel Svonava, CEO at Superlinked Demetrios Brinkmann, Agentic AI Foundation 📅 Thursday 19 November | 9:00–10:00 AM PT If you're building AI for enterprise environments and thinking seriously about security, sovereignty and inference costs, join us. 👉 Register here: luma.com/n7l3apll

Post attachment
Daniel Svonava avatar
⟳ Superlinked retweeted
Daniel Svonava@svonava·Nov 12

Heading to @sgl_project summit? Catch my talk tomorrow on building task-specific Recursive Language Models 👇 I have a love-hate relationship with these things - it’s such a cool idea but very hard to get right / working properly.

Post attachment
Superlinked avatar
Superlinked@superlinked·Nov 12

We're at the SGLang Summit 2026 tomorrow! @svonava is delivering a talk: "Build a Task-Specific RLM with SIE." Why use a large, general-purpose model for every task? Daniel will explore how to build task-specific Recursive Language Models using Superlinked Inference Engine (SIE), with a focus on making AI inference more efficient and cost-effective. 📅 **Friday, 13 November | 11:10–11:25 a.m. PT** 📍 **Decode Room, Gateway Pavilion, Fort Mason, SF** If you're attending, come along and say hello! 🔗 sglang.io/summit

Post attachment
Daniel Svonava avatar
⟳ Superlinked retweeted
Daniel Svonava@svonava·Nov 11

Check previous runs to see what went wrong before fine-tuning your model. Did the model get what it needed to begin with? At Superlinked, we build infrastructure for small models, and this distinction is important to us. For a model to prove it can do its job, the environment around it must provide a fair test :D So, start with a failing run. - Did retrieval return the required evidence? - Did the tool schema suggest what the required inputs are? - Did the retry change the outcome or did it just increase the cost? Then, repeat the same cases with a smaller model, keeping everything else (prompts, tools, test cases) the same… If both models fail at the same step, then inspect that step. If only the smaller model fails with the right inputs, it's then safe to say the model has limitations. That's why I'd first define what each tool expects as input and what it returns. If these rules keep changing, the training examples can easily stop matching the task. Once the interface is stable enough and the remaining failures point to the model, fine-tuning becomes worth testing. Measure task completion along with latency and cost, including retries. You can read the full article here: superlinked.com/blog/agent-s…

Superlinked avatar
Superlinked@superlinked·Nov 10

Can smaller AI models handle real legal work? For many repetitive, well-defined tasks, they can. Document classification, information extraction, clause retrieval and sensitive-data detection don't necessarily require a frontier LLM. The challenge is knowing which tasks to automate, how to evaluate model performance and how to combine specialised models into a reliable legal workflow. We explored these questions with Milo and Marianna, the CEO and CTO of @Fellow_legal, an AI-native law firm serving startups and investors. They joined Daniel to discuss how they're automating legal work, why off-the-shelf AI tools have limitations, and where self-hosted, task-specific models fit into their plans. Pick out the models you need for your legal tasks and run them behind one API here: auth.superlinked.com/sign-up

Post attachment
Daniel Svonava avatar
⟳ Superlinked retweeted
Daniel Svonava@svonava·Nov 7

Sovereign AI hackaton in Williamsburg def has its own vibe ^_^

Post attachmentPost attachment