Superlinked retweeted
Want to try open weight Task Specific models but don't know where to start? @superlinked has offering a grant of 500$ of credits to try them! Would love for you to check them out.
So the launch went well, we have received 265 grant applications and have many more accounts with real usage 🤗 Chart up to the right! :-)
If you are not sure if SLMs can do the job, here is a little inspiration - a summary of real projects already building on the platform 💪
0
1
2
41
Superlinked retweeted
Fast inference for Small Language Models (@svonava) is the engine behind Compute-Optimal agentic scaling: once sampling is cheap, parallel search beats raw parameter count.
Stanford CS329A: Self-Improving AI Agents · Study Notes 03 (Pass@k Power Laws & Coverage)
• Pass@k Scale math, 1 - (1 - p)^k, so a model with a 10% single-shot pass rate reaches 65% coverage in 10 tries.
• Best-of-N shifts the Compute-Optimal frontier: on the same inference budget, a 7B model sampling 64 parallel drafts outperforms a 70B model at Pass@1.
• Small Language Models win on spiky agentic workloads when high-temperature sampling produces diverse candidates instead of 64 near-identical copies.
• Verifier-Guided Search turns coverage into final accuracy, since candidate generation only helps as far as your verifier can pick the winning rollout.
Stanford CS329A: Self-Improving AI Agents · Study Notes 03 (Pass@k Power Laws & Coverage)
• Pass@k Scale math, 1 - (1 - p)^k, so a model with a 10% single-shot pass rate reaches 65% coverage in 10 tries.
• Best-of-N shifts the Compute-Optimal frontier: on the same inference budget, a 7B model sampling 64 parallel drafts outperforms a 70B model at Pass@1.
• Small Language Models win on spiky agentic workloads when high-temperature sampling produces diverse candidates instead of 64 near-identical copies.
• Verifier-Guided Search turns coverage into final accuracy, since candidate generation only helps as far as your verifier can pick the winning rollout.
🚨 Huge news 🚨 today we are launching SIE Serverless - a managed inference product focused on small models. It’s useful for spiky workloads on 20+ task specific workloads across Search, OCR and Agentic execution.
Sign up for a $500 credits launch grant 👇
It’s powered by our ngine that has received thousands of github stars in the last few months, battle-tested across finance, legal & tech. We believe that self-hosting is going to be critical given all the trust & security issues in AI, but practically there are always workloads where you’d like somebody else to deal with the GPU quota like synthetic data generation, evals, public research and beyond.
The team at Superlinked has been burning the midnight oil for months and I’m so proud of every single one of you - what a privilege to call you my team mates y’all!
We built it, now help us make some noise 🙌
A like, a comment or a repost would mean a lot! ❤️
1
2
6
144
Insanely good numbers...
Run the models and see the results for yourself: auth.superlinked.com/sign-up
Run the models and see the results for yourself: auth.superlinked.com/sign-up
"YOU NEED A BIG BOLD CLAIM IN YOUR LAUNCH"
So anyway, the evals we used for this and how robust are the findings, see: superlinked.com/blog/launch
0
0
2
170
Superlinked retweeted
"YOU NEED A BIG BOLD CLAIM IN YOUR LAUNCH"
So anyway, the evals we used for this and how robust are the findings, see: superlinked.com/blog/launch
So anyway, the evals we used for this and how robust are the findings, see: superlinked.com/blog/launch
1
2
6
454
Superlinked retweeted
Your AI agent might be using a $100 solution for a 10-cent problem.
Not every task needs your most powerful model.
Classification, OCR, extraction, reranking and guardrails can often run on smaller, more cost-efficient models.
That's the problem Superlinked SIE is solving.
behind your agent.
Keep frontier intelligence for frontier work.
Not every task needs your most powerful model.
Classification, OCR, extraction, reranking and guardrails can often run on smaller, more cost-efficient models.
That's the problem Superlinked SIE is solving.
behind your agent.
Keep frontier intelligence for frontier work.
🚨 Huge news 🚨 today we are launching SIE Serverless - a managed inference product focused on small models. It’s useful for spiky workloads on 20+ task specific workloads across Search, OCR and Agentic execution.
Sign up for a $500 credits launch grant 👇
It’s powered by our ngine that has received thousands of github stars in the last few months, battle-tested across finance, legal & tech. We believe that self-hosting is going to be critical given all the trust & security issues in AI, but practically there are always workloads where you’d like somebody else to deal with the GPU quota like synthetic data generation, evals, public research and beyond.
The team at Superlinked has been burning the midnight oil for months and I’m so proud of every single one of you - what a privilege to call you my team mates y’all!
We built it, now help us make some noise 🙌
A like, a comment or a repost would mean a lot! ❤️
3
67
292
48,401
Superlinked retweeted
Each agent startup after seeing the inference invoice:
🚨 Huge news 🚨 today we are launching SIE Serverless - a managed inference product focused on small models. It’s useful for spiky workloads on 20+ task specific workloads across Search, OCR and Agentic execution.
Sign up for a $500 credits launch grant 👇
It’s powered by our ngine that has received thousands of github stars in the last few months, battle-tested across finance, legal & tech. We believe that self-hosting is going to be critical given all the trust & security issues in AI, but practically there are always workloads where you’d like somebody else to deal with the GPU quota like synthetic data generation, evals, public research and beyond.
The team at Superlinked has been burning the midnight oil for months and I’m so proud of every single one of you - what a privilege to call you my team mates y’all!
We built it, now help us make some noise 🙌
A like, a comment or a repost would mean a lot! ❤️
0
2
109
129,089
Superlinked retweeted
bills will continue until morale improves (or small models get deployed) ^_^
0
1
2
47
Superlinked retweeted
SuperExcited for @superlinked 's launch! For the past month, I have been 'in-residency' at Superlinked during my sabbatical. Here's the deets!
a) You don't want to always want to be managing the model containers in production, bringing them up and tearing them down with simple yaml configurations. You want a battle tested engine that can autoscale the nodes based on what's actively being asked and load balance the requests esp. when they take a lot of time serving tokens for LLMs.
b) Not everything has to be an LLM call! Specialized models are thriving and you can use one of the many open-weights models that are task specific to do the same work faster, cheaper and more effectively. Wrap them in tool calls or make them part of your deterministic workflows!
c) Running models on your own infrastructure costs a lot! SIE helps reduce your spend by having model eviction and fractionally using the GPUs by putting multiple SLMs on a single one. This allows for a much more leaner stack. With SIE serverless in the cloud, you get the same benefits and allows you to try these models before you start adapting these and serving them on your own infrastructure.
The residency is allowing me a look into how one of the sharpest teams runs production grade inference models. At the same time, I get to talk to developers and contribute to the Open Source AI ecosystem. I'm collaborating on hosting a few events in NY, SF, Online + attending Kubecon, and more! Looking forward to the chat!
a) You don't want to always want to be managing the model containers in production, bringing them up and tearing them down with simple yaml configurations. You want a battle tested engine that can autoscale the nodes based on what's actively being asked and load balance the requests esp. when they take a lot of time serving tokens for LLMs.
b) Not everything has to be an LLM call! Specialized models are thriving and you can use one of the many open-weights models that are task specific to do the same work faster, cheaper and more effectively. Wrap them in tool calls or make them part of your deterministic workflows!
c) Running models on your own infrastructure costs a lot! SIE helps reduce your spend by having model eviction and fractionally using the GPUs by putting multiple SLMs on a single one. This allows for a much more leaner stack. With SIE serverless in the cloud, you get the same benefits and allows you to try these models before you start adapting these and serving them on your own infrastructure.
The residency is allowing me a look into how one of the sharpest teams runs production grade inference models. At the same time, I get to talk to developers and contribute to the Open Source AI ecosystem. I'm collaborating on hosting a few events in NY, SF, Online + attending Kubecon, and more! Looking forward to the chat!
🚨 Huge news 🚨 today we are launching SIE Serverless - a managed inference product focused on small models. It’s useful for spiky workloads on 20+ task specific workloads across Search, OCR and Agentic execution.
Sign up for a $500 credits launch grant 👇
It’s powered by our ngine that has received thousands of github stars in the last few months, battle-tested across finance, legal & tech. We believe that self-hosting is going to be critical given all the trust & security issues in AI, but practically there are always workloads where you’d like somebody else to deal with the GPU quota like synthetic data generation, evals, public research and beyond.
The team at Superlinked has been burning the midnight oil for months and I’m so proud of every single one of you - what a privilege to call you my team mates y’all!
We built it, now help us make some noise 🙌
A like, a comment or a repost would mean a lot! ❤️
2
1
7
257
Superlinked retweeted
Stop overpaying for agents, folks!
A single agent request can trigger many model calls. Your agent might classify a ticket, rerank results, extract a date and run a guardrail before it ever answers.
Sending all of those calls to a frontier model gets expensive fast. And it’s unnecessary.
@superlinked SIE gives you one API for the open models behind those steps.
Across 14 task tests, open models on SIE reached a median 98% of GPT-6.1 Sol's score at 3% of its cost, at list pricing.
Keep the frontier model for the steps that need it, like those requiring complex reasoning.
A single agent request can trigger many model calls. Your agent might classify a ticket, rerank results, extract a date and run a guardrail before it ever answers.
Sending all of those calls to a frontier model gets expensive fast. And it’s unnecessary.
@superlinked SIE gives you one API for the open models behind those steps.
Across 14 task tests, open models on SIE reached a median 98% of GPT-6.1 Sol's score at 3% of its cost, at list pricing.
Keep the frontier model for the steps that need it, like those requiring complex reasoning.
🚨 Huge news 🚨 today we are launching SIE Serverless - a managed inference product focused on small models. It’s useful for spiky workloads on 20+ task specific workloads across Search, OCR and Agentic execution.
Sign up for a $500 credits launch grant 👇
It’s powered by our ngine that has received thousands of github stars in the last few months, battle-tested across finance, legal & tech. We believe that self-hosting is going to be critical given all the trust & security issues in AI, but practically there are always workloads where you’d like somebody else to deal with the GPU quota like synthetic data generation, evals, public research and beyond.
The team at Superlinked has been burning the midnight oil for months and I’m so proud of every single one of you - what a privilege to call you my team mates y’all!
We built it, now help us make some noise 🙌
A like, a comment or a repost would mean a lot! ❤️
17
4
38
8,536
Superlinked retweeted
Signup & apply for the grant (tell us what you are building to help us prioritize): superlinked.com
2
3
15
13,799
Superlinked retweeted
using gpt for "is this a refund request y/n" is like taking a plane to the corner store
2
1
2
149
Superlinked retweeted
just saw this and i'm still processing it.
text classification on GPT-6.1 Sol: 1.48 seconds.
same task on SIE: 13.1 milliseconds.
that's over 100x faster.
and your agent doesn't do it once. it does it on every single request.
Superlinked just open-sourced the fix: one API the right model and saves the big one for the hard stuff.
$20 free. no card.
text classification on GPT-6.1 Sol: 1.48 seconds.
same task on SIE: 13.1 milliseconds.
that's over 100x faster.
and your agent doesn't do it once. it does it on every single request.
Superlinked just open-sourced the fix: one API the right model and saves the big one for the hard stuff.
$20 free. no card.
🚨 Huge news 🚨 today we are launching SIE Serverless - a managed inference product focused on small models. It’s useful for spiky workloads on 20+ task specific workloads across Search, OCR and Agentic execution.
Sign up for a $500 credits launch grant 👇
It’s powered by our ngine that has received thousands of github stars in the last few months, battle-tested across finance, legal & tech. We believe that self-hosting is going to be critical given all the trust & security issues in AI, but practically there are always workloads where you’d like somebody else to deal with the GPU quota like synthetic data generation, evals, public research and beyond.
The team at Superlinked has been burning the midnight oil for months and I’m so proud of every single one of you - what a privilege to call you my team mates y’all!
We built it, now help us make some noise 🙌
A like, a comment or a repost would mean a lot! ❤️
10
3
86
29,186
We're excited that you're excited, thanks for the shoutout @kairosdojo !
0
0
1
14
Superlinked retweeted
NOBODY TALKS ABOUT THIS AND IT'S THE BIGGEST BILL IN YOUR AGENT
it's not the llm you picked.
it's that every step of the workflow (classify, rerank, extract, run a guardrail) goes to the same frontier model because it's the easy option.
at scale, most of your inference bill is for work a much smaller model would handle just as well.
Superlinked ships hosted SIE today: one api to pick the right model for each task without rewriting anything.
open source, Apache 2.0, 3.4k stars on github.
repo link below.
it's not the llm you picked.
it's that every step of the workflow (classify, rerank, extract, run a guardrail) goes to the same frontier model because it's the easy option.
at scale, most of your inference bill is for work a much smaller model would handle just as well.
Superlinked ships hosted SIE today: one api to pick the right model for each task without rewriting anything.
open source, Apache 2.0, 3.4k stars on github.
repo link below.
🚨 Huge news 🚨 today we are launching SIE Serverless - a managed inference product focused on small models. It’s useful for spiky workloads on 20+ task specific workloads across Search, OCR and Agentic execution.
Sign up for a $500 credits launch grant 👇
It’s powered by our ngine that has received thousands of github stars in the last few months, battle-tested across finance, legal & tech. We believe that self-hosting is going to be critical given all the trust & security issues in AI, but practically there are always workloads where you’d like somebody else to deal with the GPU quota like synthetic data generation, evals, public research and beyond.
The team at Superlinked has been burning the midnight oil for months and I’m so proud of every single one of you - what a privilege to call you my team mates y’all!
We built it, now help us make some noise 🙌
A like, a comment or a repost would mean a lot! ❤️
14
6
171
42,332
