Enrique J. Cardona

Co-founder & CTO @Finanboo 🎋 // Lead Developer @H2Stdio. 📚 Learning everyday something new...

[Enrique J. Cardona @henry2man]

@lmontoya @jerolba Exacto, es que es tan simple como tener tu inferencia local como "fuerza bruta", y luego, por ejemplo, una suscripción de 20€ de Cursor para poder afinar, hacer diagnósticos y poder consultar a los modelos grandes.

[Enrique J. Cardona @henry2man]

@jerolba Lo que no comprendo es que le pongan precio similar al DGX Spark, pero sin incluir la red de alta velocidad de 200Gbps, cuando solo el hardware de red está valorado en unos 1000-1500$.

A priori, en igualdad de condiciones, para mi es un NO-GO de libro :/

[Bernardo Quintero @bquintero]

Europa se preocupó mucho por los riesgos de los modelos avanzados de IA.

Quizá no lo suficiente por los riesgos de no tenerlos.

[Jun Song @jun_song]

Now, countries outside the US and China need to wake up to the urgency of Sovereign AI and start building.

But building a proprietary LLM from scratch means hitting a wall—they lack everything from crucial training data to funding and compute.

The move they need to make is to adopt a strategy like Cursor Composer.

They need to focus entirely on the post-training of open-weight models.

Look at Cursor. They took Kimi-K2.5—a model from two generations ago—and turned it into an Opus-level model purely through post-training, radically cutting costs.

I believe this post-training market is going to absolutely explode this year.

[Enrique J. Cardona @henry2man]

@DotCSV 💥

[Samuel Solís @estoyausente]

No puede depender nuestro trabajo/empresa de estas cosas.

El riesgo es altísimo.

Puedes ser una empresa IA-native si quieres, pero no puedes depender de otra tan claramente. Si no tienes tus propias IA s autoalojadas, estás a un movimiento así del cierre.

[Enrique J. Cardona @henry2man]

@root_rat Pues yo ya tengo mi segundo Nvidia GB10 en camino…

[0xSero @0xSero]

State of Local AI #1

———

In lieu of Fable ban.

Here’s the best LLMs of the week to run on your hardware.

—— 4-8gb vram/ram 500$

- Gemma-4-qat https://t.co/UFCmLXVKed I had someone mention it’s very good for subagent stuff

—— 8-16gb vram/ram < 1k usd

- Gemma-12B https://t.co/tc6IBTrbc3 without a doubt the smartest model of its size

—— 16-32gb Apple/Strix halo 1-2k usd

- Diffusion Gemma26B https://t.co/mSaWPFpgXQ

- on 1x 6000 it’s eating up to 600 tok/s
- smallest smart MoE we have
- lots of world knowledge
- easy to run

—— 32-96gb ram/vram (2-10k usd)

- nex-n2-mini https://t.co/EL1ePzwI58 builds on qwen3.6-35B and seems to do really well

- qwopus-27B https://t.co/P1gypZwufi this model topped a lot of our benchmarks at https://t.co/UfoYoOlSIk

—— 384gb vram (10-50K usd)

- https://t.co/AZb0Gtu5P3 23B means it’s close to qwen3.6-27B per token, while also have a lot of specialisation.

- fast inference
- top open weight model on AA

—— 768gb-1TB

- https://t.co/kWzJG2Hjen

Kimi has always been a top player here and their last model cuts speed and cost down by 30%
- great vision support
- first coder model by moonshot

———

Top models:

1. Qwen3.6-35B
2. Qwen3.6-27B
3. Step-3.7-Flash
4. Minimax-M3
5. Deepseek-v4-flash

———

Budget sweet spots:

#1 - 1K usd

Single 3090 / Mac mini / Intel arc b70 / AMD

- Qwen / Gemma

#2 - 5k usd

DGX Spark / Mac m5 max / 4x 3090

- qwen / Gemma step and deepseek flash

#3 - 12k usd

RTX Pro 6000 / Mac Ultra / 2x Spark / 8x 3090

Ds4-flash / step-3.7-Flash and above

#4 - 24k usd

2x 6000 / 2x Mac Ultra / 4x Spark / Mix

Same as above

#5 - 50k usd

4x 6000 / 4x Max Ultra / 12x Spark / 2 H100

Minimax-m3 / nex-n2-pro / step-3.7-flash

#6 - 100k usd

GB300 station / 8x 6000 / 4x H200 / Mix

GLM-5.2 / Kimi-K2.7

———

Let’s keep the Internet free thanks for reading

[Enrique J. Cardona @henry2man]

@mkurman88 @TheAhmadOsman Ahmad one year later:

[Enrique J. Cardona @henry2man]

.@TheAhmadOsman Just a friendly reminder...

#Fable #Mithos #Anthropic

[Ahmad @TheAhmadOsman]

friendly reminder to buy a GPU and secure your compute on this wonderful evening

your AI cannot be controlled by a self-serving corporate

[Enrique J. Cardona @henry2man]

@root_rat Yo ya llevo tiempo animando a todo el mundo a aprender con lo que tenga. Afortunadamente ya existen modelos pequeños pero matones que se pueden ejecutar en máquinas modestas.

Y además muchísimas tareas no necesitan un modelo gigante para resolverse.

This is the way!

[Enrique J. Cardona @henry2man]

Today, an more than ever

*Open Source must win*

[0xSero @0xSero]

https://t.co/txDnDa0Flf

[Finanboo @finanboo]

El BCE congela tipos al 2,25%.

¿Sabes cuánto te cuesta tu deuda bancaria este mes? No "más o menos". El número exacto.

La mayoría de las pymes no lo tienen claro. Y eso tiene un coste concreto cuando negocian con el banco.

#BCE #pymes #gestionfinanciera

[sparkarena @spark_arena]

day 0 support on sparkrun:

sparkrun update

sparkrun run @eugr/diffusion-gemma-bf16-thinking sparkrun run @eugr/diffusion-gemma-nvfp4-thinking sparkrun run @eugr/diffusion-gemma-nvfp4

Check sparkrun list for other options

[sparkarena @spark_arena]

@UnslothAI day 0 support on sparkrun:

sparkrun update

sparkrun run @eugr/diffusion-gemma-bf16-thinking sparkrun run @eugr/diffusion-gemma-nvfp4-thinking sparkrun run @eugr/diffusion-gemma-nvfp4

Check sparkrun list for other options

[FullStack Sevilla @fullstackSVQ]

Tenemos el último meetup de la temporada en #FullstackSevilla.

Esta vez cambiamos la IA por otro tema que tarde o temprano llega a muchos developers: el liderazgo.

Junto a @RafaSobra hablaremos de los retos de pasar de "solo picar código" a ayudar a que un equipo tenga éxito.

[FullStack Sevilla @fullstackSVQ]

Da igual si te llaman Tech Lead, Engineering Manager, Senior Engineer o cualquier otra cosa.

Si te toca tomar decisiones, ayudar a otros o liderar iniciativas, este meetup es para ti.

💬 Charla práctica
❓Preguntas
🍻 Networking
¡Nos vemos mañana en #FullstackSevilla! 🚀

[blue @bluewmist]

I'm finally reading Dune. This quote, which is in the first few pages, hits hard:

"Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them."

[Enrique J. Cardona @henry2man]

@0xSero Same results as @mikhei777 :/

Seems too big for the spark. OOMs everywhere.

Dual Spark will be the only reliable way, I think.

[Xuan-Son Nguyen @ngxson]

llama.cpp is likely the first LLM runtime in the world to allow "interrupt" reasoning without stopping the whole response.

We also added a small "skip" button on the Web UI, the model gives the final response as soon as you click the button.

The response is no longer bound to reasoning budget!