ra

@barckcode @danielvigueras Al final has escrito ese post detallado? Gracias

@danielvigueras Con vLLM y LiteLLM. Sacaré un post más detallado de cómo he montado todo

Un último dato curioso. En ~3 días que lleva el server de inferencia up. Hemos quemado ya +mil millones de tokens 😂 Si esto fuese una API de algún servicio privado, depende de qué provider ya nos habría rentabilizado el coste del servidor incluso jaja Que barbaridad

@ErickSky @UnslothAI No pasa de 5-8 token/s haciendo offloading casi seguro.

💣 Si eres developer, builder de agentes, automatizas workflows, haces devops, frontend heavy o simplemente te encanta tener el control total de tu AI… este es el anuncio que estabas esperando. @UnslothAI lo acaba de hacer realidad el sueño de tener Kimi K2.6 de Moonshot AI, un modelo de 1 TRILLÓN de parámetros corriendo LOCALMENTE en tu PC… y a más de 40 tokens/segundo. El monstruo open-source que estaba rompiendo todos los benchmarks de coding y agentes ya fue cuantizado con su nueva tecnología Dynamic GGUFs v2.0. Algunos datos sobre Kimi K2.6: - Líder open-source en coding: SWE-Bench Verified 76.8, Multilingual 73.0, BrowseComp 74.9, DeepSearchQA 77.1… - Agentes de largo plazo reales: más de 4.000 tool calls seguidos y 12+ horas de ejecución sin parar. - Soporta de todo: Rust, Go, Python, frontend con videos, WebGL, shaders, Framer Motion, Three.js… - Swarm de agentes nivel dios: hasta 300 sub-agentes paralelos trabajando al mismo tiempo. - Agentes proactivos 24/7, Claw Groups (traes tus propios agents + humanos + bots en loop infinito)… Usando Dynamic GGUFs v2.0 no cuantizaron todo por igual (como hacen los demás). Upcastean dinámicamente las capas clave de razonamiento y coding, mientras comprimen el resto. Resultado: mantienes casi toda la inteligencia del modelo full-precision… pero ahora pesa solo 340GB. Las specs: - 340GB → más de 40 tokens/segundo en configuraciones de ~350GB (CPU/GPU/SSD) - Versión full precision: 610GB - Corre en CPU, GPU o incluso streaming desde SSD El open-source oficialmente ha alcanzado (y superado) al closed-source. ENLACES👇

@llmdevguy @UnslothAI @Kimi_Moonshot Would you give some feedback if you test this configuration? Thanks 🙏

@UnslothAI @Kimi_Moonshot So if I would have rtx 6000 pro + 256gb ram, it should run around 40 t/s?

Kimi K2.6 can now run on CPU, GPU and SSD setups! 🔥 We shrank the 1T model to 340GB via Dynamic GGUFs where important layers are upcasted. Run at >40 tok/s on 350GB RAM/VRAM setups. Run full precision on 610 GB. Guide: unsloth.ai/docs/models/ki… GGUF: huggingface.co/unsloth/Kimi-K…

Are you THIS OLD 😅?

@geekdegafas Wifi mesh 3 repetidores

Hoy y mañana hay ofertazas en Amazon. Pero hay que bucear un poco para despejar la morralla. Si buscas algo concreto, dime y te pongo la mejor opción (en oferta).

Madre mía lo que encontré ayer en casa @migueldeicaza

@Hesamation @maximelabonne Any organized resource after this?

a reminder that there is a perfect free LLM course on github with fully-detailed roadmaps for LLM scientist and engineer careers. it has notebooks, resources, papers, and everything you need as a beginner. stop looking for the perfect roadmap, stick to this.

@skalskip92 I mean if pose points have associated a person id that is consistent ( same person, same id) during the scene. I usually have done this in 2 steps, tracking people, pose estimation and post processing iou of both outputs to pair person id with pose.

@skalskip92 Does it track people id?

ViTPose++ is crazy good! look at the interaction between pink and green player. getting that right is really impressive. we are plugging it into basketball AI.

We're thrilled to announce SignGemma, our most capable model for translating sign language into spoken text. 🧏 This open model is coming to the Gemma model family later this year, opening up new possibilities for inclusive tech. Share your feedback and interest in early testing → goo.gle/SignGemma

@reach_vb RAM needed? Too much for sure.

Optimised MLX quant for DeepSeek R1 0528 🔥

@javisantana Lo haces. En las carreras sale ese puntito más

So my plans was to run a 10km race in 50 minutes on May-31st. This one is special because it takes place in the same race track I usually go with my car (see my bio) current status (from an easy run, aerobic) I did 49:00 in a 9.6km race a few days ago but seems hard I can go 3 minutes down in 15 days

Join us for the third MULTIDATA webinar: a hands-on intro to our project and pipeline!🗓️ May 19, 4 PM (CET)🎤 Cristóbal Pagán (project coordinator) 🔍 Learn about speech transcription, gesture tracking, and more — no tech skills needed! 🎟️ Save your spot: …roduction-to-MULTIDATA.eventbrite.com

Join us for the MULTIDATA Summer Workshop on Tuesday, July 8 at the Max Planck Institute (Nijmegen) — right before the ISGS conference! Free registration is open now: …ta-summerworkshop-2025.eventbrite.com

We made a Guide to teach you how to Fine-tune LLMs correctly! Learn about: • Choosing the right parameters & training method • RL, GRPO, DPO & CPT • Data prep, Overfitting & Evaluation • Training with Unsloth & deploy on vLLM, Ollama, Open WebUI 🔗 docs.unsloth.ai/get-started/fi…

❤️

For years, I was hyperparameter tuning XGBoost models wrong. In 3 minutes, I'll share one secret that took me 3 years to figure out. When I did, it cut my training time 10X (and gave an instant 25% boost in model accuracy). Let's dive in. 🧵

HOLY SHIT thought-to-action is almost here meta just published two papers showing they can decode thoughts into text with 80% accuracy, in real time, using non-invasive brain-computer interfaces. pardon my french, but WHAT THE FUCK?

"We aim to mobilise a total of €200 billion for AI investments in Europe. It will be the largest public-private partnership in the world for the development of trustworthy AI." — President @vonderleyen at the Artificial Intelligence Action Summit

What if a [MASK] was all you needed? ModernBERT is great, but we couldn't stop wondering if it could be greater than previous encoders in different ways. Maybe we don't need task-specific heads? Maybe it can do all sort of tasks with only its generative head? Spoilers: Yes