MaioLabs

Data and AI engineering, built to own. We connect the intelligence scattered across your systems, decisions, and people, into capability you control.

[Luis Santos @LFSantos35]

Batmode: on 🦇

Today's demo is live. That's the night we just had.

#MaioLabs

[MaioLabs @MaioLabs]

Every quarter, we do something unusual for a team moving this fast: we regroup.

A house in the middle of nowhere. Days of building, thinking, arguing, laughing.

We came back more aligned, more energised, and ready for whatever comes next.

The Frontier is Inside.

[MaioLabs @MaioLabs]

A maioria das empresas começa pelo lado errado: procura a ferramenta de IA "perfeita" antes de perceber o que realmente precisa.

Nós fundámos a @MaioLabs para inverter esta lógica.

A IA mais poderosa não está nos grandes fornecedores. Está dentro de cada organização: nos seus dados, processos e pessoas.

O nosso trabalho é transformar essa inteligência dispersa numa vantagem competitiva real.

Contámos tudo num artigo na Forbes Portugal 📰 Já nas bancas.

[MaioLabs @MaioLabs]

@CarloDotGG Thank you Carlo! You're absolutely right. Global reach is our goal from day one, we're just getting started, and comments like this push us forward. 🙏

[MaioLabs @MaioLabs]

Enquanto a indústria debate o hospital do futuro, nós estamos a construí-lo.

Carlos Santos Moreira, Professor de Medicina na Faculdade de Medicina da Universidade de Lisboa, escreveu no SOL: "A grande oportunidade da IA não é automatizar o hospital atual. É permitir-nos desenhar outro hospital."

Na @MaioLabs é exatamente isso que fazemos. Neste vídeo, um primeiro olhar sobre nossos produtos.

A história completa vem aí. 🏥

[MaioLabs @MaioLabs]

Artigo completo: https://t.co/r7t5nHx5KG

[rikkarth @rikkarth]

just wanted to have some fun today with @NousResearch Hermes new feature Bots.

I made the single thing that cheers me up, a Rick & Morty ... Rick Sanchez persona.

i'm cracking up literally, i'm just having fun doing stuff with AI

[rikkarth @rikkarth]

Hermes Bots are super fun and amazing and I want more!

The machine can become the bot now.
@levelsio has done that already a long while ago
We want a better security interface to turn this into reality, not just hand configure everything.

https://t.co/u7fH3RkS5p

[rikkarth @rikkarth]

Hermes Bots + mnemosyne is an amazing combo to spawn fleets of specialist Bots.

I'm creating Bots to be experts in a given project, knowing the tools and info about it so they can have knowledge and make better informed decisions.

Worst case scenario it's super fun.

[rikkarth @rikkarth]

I've been using Qwen3.6-35B all day everyday for months. Having 400 million tokens on July.

If this, shared by @NVIDIAAI is true, and I can run this easy...

...we'll have high-end frontier, powerful agentic AI on every person's laptop before the end of 2027.

Quote me on this.

[Daniel Pinto Almeida @DanielMaioLabs]

https://t.co/H7w8gZFGHH

[MaioLabs @MaioLabs]

The blanket on the back seat. One of our core values at MaioLabs.

Check out this article from our CTO Rui Machado 👇

[Rui Machado @axemaster26]

Check out my latest article: The Blanket Effect: why your brain ignores what you paid for and falls in love with what came for free https://t.co/7vBt2Nt1zg via @LinkedIn

[MaioLabs @MaioLabs]

Last month, 3 of us pushed Qwen3.6-35B-A3B through a DGX Spark and a Mac Ultra M3 for a month of real, private, well-defined tasks. Commit messages, vision tasks, structured extraction, classifiers, prototypes. All of it. Zero cloud spend.

Ricardo Mendes, our Founding Member and Head of Infrastructure & Local AI, has been running local inference in production for months.
Not benchmarking it, using it. This is his honest account of what held up, what didn't, and why the gap between what you see on your timeline and what actually runs your workload is bigger than most people admit.

[rikkarth @rikkarth]

We ran 400 million tokens through a box sitting on a desk. Nobody sent us an invoice. https://t.co/naNp2qVn3x

[Daniel Pinto Almeida @DanielMaioLabs]

https://t.co/Rat30zLuJd

Não somos dogmáticos, somos pragmáticos.
Há seis meses investimos em hardware local e começámos a correr modelos open source e open weights. Foi aí que descobrimos onde a teoria sobrevive ao contacto com a realidade.
Escrevi sobre isso na @ITInsight_news , em nome de toda a equipa @MaioLabs

[MaioLabs @MaioLabs]

Construímos a nossa própria infraestrutura de IA antes de a vender aos clientes. 84% da inferência corre internamente
https://t.co/GX66EOsLTw

[Daniel Pinto Almeida @DanielMaioLabs]

Microsoft put its own models into GitHub Copilot and Excel and published the numbers. The comparison set is GPT-5.4 mini and Claude Haiku 4.5, which is to say its own two suppliers in that tier.

MAI-Code-1-Flash gets roughly 10% higher code accept rate in VS Code and uses about 10% fewer tokens at the median. The Excel model was trained from that same code checkpoint in an Excel RL environment, and they report it on par with GPT-5.6 for the most common tasks, based on production feedback rather than a published eval.

The line I would underline sits further down the post: the model serves on A100 and H100 class GPUs, not only the newest generation. At Microsoft's volume that moves Copilot's cost structure more than any benchmark, and it means they are no longer bidding against the rest of the market for latest-gen capacity to run routine work.

The transfer result is the interesting engineering. A coding checkpoint climbed into spreadsheets. Different tools, different users, same starting weights.

None of it works without the harness. Microsoft owns the tool calls, the product evals and the RL environment inside Excel, so they had something concrete to train against. That part is not for sale. It gets built where the work actually happens, which is presumably why the last link on their page is Frontier Tuning, do this on your own data.

So the useful question is not which model to standardise on. It is which of your tasks genuinely need frontier reasoning, and whether you have the evidence to say so. Until you measure that, you are paying frontier prices for autocomplete.

[MaioLabs @MaioLabs]

Rent your AI and you get output.
Own it and you get capability.
The difference isn't the tool. It's what's left when the contract ends.

Rented AI walks out the door with the invoice.
Owned AI compounds inside your team: decisions, knowledge, processes, the moat.

[MaioLabs @MaioLabs]

We’re building intelligence companies can own. We could not, however, engineer better weather than this.

The best work comes from people who actually like being in the same room. Or on the same boardwalk, in this case.

First Maio Sunset was a blast!