Enrique J. Cardona

@alexocheema @dbarkalov @exolabs @nvidia @alexocheema What's the realistic schedule for DGX Spark GB10 support? I have a Mac Studio that needs a new friend.

@dbarkalov @exolabs @nvidia Releasing this soon with full support in exo

Clustering NVIDIA DGX Spark + M3 Ultra Mac Studio for 4x faster LLM inference. DGX Spark: 128GB @ 273GB/s, 100 TFLOPS (fp16), $3,999 M3 Ultra: 256GB @ 819GB/s, 26 TFLOPS (fp16), $5,599 The DGX Spark has 3x less memory bandwidth than the M3 Ultra but 4x more FLOPS. By running compute-bound prefill on the DGX Spark, memory-bound decode on the M3 Ultra, and streaming the KV cache over 10GbE, we are able to get the best of both hardware with massive speedups. Short explanation in this thread & link to full blog post below.

It's hard to believe that the "~80%+ of the internet is blocked in Spain during football games" claim is true - but it is!! And has been for years. The government is sabotaging their complete digital economy... for La Liga, a private football org worth €5B. Pure madness

😡 Un ilustrador catalán. 38 años cotizando como autónomo. Sin fallar un solo mes. En abril de 2025 hace los trámites de jubilación. Han pasado 12 meses. No cobra. Esto es lo que le ha pasado: Presenta la solicitud en la Seguridad Social de Terrassa. Le dicen que todo está correcto. Pasan tres meses sin respuesta. Vuelve a preguntar. Le informan de que su expediente se ha enviado a Cádiz. Nunca ha cotizado en Cádiz. Ni fuera de Cataluña. La explicación: "estaban saturados de trabajo." El expediente llega a Cádiz. Desde allí dicen que han encontrado "una irregularidad en 1988 y 1989." Devuelven el expediente a Cataluña. Cataluña lo revisa, ve que todo es correcto, y lo reenvía a Cádiz. Cádiz lo devuelve exigiendo que se explique "por qué es correcto." La pelota cambia de tejado tres veces. Nadie resuelve. Intenta hablar con la persona que lleva su expediente en Cádiz. Le dicen que está de vacaciones. Que ya le llamarán. Pasan dos meses. Nadie llama. Contrata a un gestor laboralista. El gestor envía un escrito a la Seguridad Social. El INSS tiene 40 días para responder. No responde. Silencio administrativo. El gestor consigue localizar al funcionario que lleva el caso. Sorpresa: ya no está en Cádiz. Lo han trasladado a Sevilla. Le dicen que le "darán prioridad." No se la dan. Escribe al Síndic de Greuges. Escribe al Defensor del Pueblo. La respuesta: "El documento ha entrado." Punto. Hoy, abril de 2026, este hombre lleva un año sin cobrar la pensión que le corresponde. Más de 11.000€ que no ha recibido. Y como no tiene la jubilación aprobada, tampoco puede acceder a su propio plan de pensiones. 38 años pagando. Ni un mes sin cotizar. Y el sistema no es capaz de tramitar su jubilación en 12 meses. No le han denegado la pensión. Eso tendría solución — recurres y listo. Lo que le hacen es peor: no le dicen ni sí ni no. Le marean. Le pasean el expediente entre tres ciudades. Le dejan en un limbo donde no trabaja, no cobra y no puede reclamar sin ir a juicio. Y si va a juicio, puede tardar otro año. Y las costas las paga él. Esto no es un caso aislado. Es el sistema funcionando como funciona: sin responsabilidad individual, sin plazos reales, sin consecuencias para quien no responde. Si le haces esto a un cliente, te denuncian. Si te lo hace la Administración, te dicen que "el documento ha entrado."

What a sight!!! 😍

EARTHSET. April 6, 2026. Humanity, from the other side. First photo from the far side of the Moon. Captured from Orion as Earth dips beyond the lunar horizon. Photo: NASA

guess who released yet another fine tune? me! here is a qlora for Qwen3.5 35b a3b MoE which uses the same dataset and technics than jackrong/qwopus3.5v3 with slight changes to fit with the moe structure. benchmarks hold well. available in q8, q6, q5 and q4. thanks @johnny_everson for the suggestion! huggingface.co/samuelcardillo…

Memory bandwidth for local AI hardware matters a lot more than most people think People keep comparing boxes like this: model size vs memory capacity That is only half the story The better mental model is: > capacity = what fits > bandwidth = how hard it can breathe > software stack = how much of that you actually cash out You are buying a memory subsystem and then negotiating with physics Here is the current local AI hardware ladder: > RTX PRO 6000 Blackwell > 96GB > 1792 GB/s > RTX 5090 > 32GB > 1792 GB/s > RTX 4090 > 24GB > 1008 GB/s Raw single-card bandwidth king stuff Now Apple > Mac Studio M3 Ultra > up to 512GB unified memory > 819 GB/s > Mac Studio M4 Max > up to 128GB > 546 GB/s > MacBook Pro M5 Max > up to 128GB > 460 to 614 GB/s > MacBook Pro M5 Pro > up to 64GB > 307 GB/s > Mac mini M4 Pro > up to 64GB > 273 GB/s > MacBook Air M5 > up to 32GB > 153 GB/s Apple is not winning raw bandwidth vs top NVIDIA Apple is winning the: > “I want one quiet box with a stupid amount of usable memory” argument And that is still a very real argument Now another interesting new category > DGX Spark > 128GB unified memory > 273 GB/s > GB10 class boxes like ASUS Ascent GX10 > 128GB unified memory > 273 GB/s These are not bandwidth monsters They are coherent-memory NVIDIA CUDA appliances That matters Because 128GB in one box changes what fits locally, even if it does not magically outrun a 5090 once the same model fits on both + CUDA Then there is the one category that actually made x86 interesting again for local AI: > Ryzen AI Max / Strix Halo > up to 128GB unified memory > 256 GB/s > up to 96GB assignable to GPU on Windows This is also where the Framework Desktop matters Not “just another mini PC” This is one of the first mainstream x86 boxes where local AI starts feeling like a serious hardware class instead of a laptop pretending very hard Then the trap people keep falling into: Most “AI PCs” are not in this tier They are down here: > Snapdragon X Elite > 135 GB/s > Intel Lunar Lake > 136 GB/s > Snapdragon X2 Elite > 152 to 228 GB/s depending on SKU > regular Ryzen AI 300 class way closer to thin-and-light territory than Strix Halo These are fine machines But the AI sticker does not create memory bandwidth Physics is still in charge which is rude but consistent AMD discrete cards > RX 7900 XTX > 24GB > 960 GB/s > Radeon PRO W7900 > 48GB > 864 GB/s > Radeon AI PRO R9700 > 32GB > 640 GB/s Not the CUDA default answer but definitely not irrelevant Intel is interesting now too > Arc Pro B65 > 32GB > 608 GB/s > Arc Pro B60 > 24GB > 456 GB/s And then there is Tenstorrent > Tenstorrent Wormhole n300 > 24GB > 576 GB/s > Tenstorrent Blackhole p150 > 32GB > 512 GB/s Not mainstream but absolutely relevant if you care about alternative and opensource local AI stacks So what does all of this actually mean? It means the local AI market is really five different markets wearing the same buzzword > fastest raw speed when it fits discrete NVIDIA > biggest one-box memory story Apple Ultra > coherent NVIDIA appliance DGX Spark / GB10 > first x86 unified-memory contender Strix Halo / Ryzen AI Max > oss stack Tenstorrent That is why people keep talking past each other A 5090 can absolutely embarrass a lot of unified-memory boxes if the model fits A Mac Studio M3 Ultra can fit things a 5090 cannot dream of fitting in one card A DGX Spark is interesting because it is compact coherent NVIDIA with 128GB & 273 GB/s + CUDA A Strix Halo box is interesting because it finally gives x86 a real answer to “what if I want big local models in one machine without going full workstation GPU?” Now Stop asking: > which box is best? Start asking: > what must fit? > what bandwidth tier do I need? > what software stack do I trust? > which bottleneck am I buying? That is how you stop guessing That is how you actually design a local AI system And yes most people still need to Buy a GPU

📣 La liberación de los datos mercantiles es una necesidad nacional. Son datos públicos, pero su modelo de explotación impide aprovecharlos justo para lo que más falta hace: detectar irregularidades en contratos públicos. ✍️ Mi tribuna en @elnotarioSXXI elnotario.es/opinion/opinio…

Me representa

There are no words.

Something I've been thinking about - I am bullish on people (empowered by AI) increasing the visibility, legibility and accountability of their governments. Historically, it is the governments that act to make society legible (e.g. "Seeing like a state" is the common reference), but with AI, society can dramatically improve its ability to do this in reverse. Government accountability has not been constrained by access (the various branches of government publish an enormous amount of data), it has been constrained by intelligence - the ability to process a lot of raw data, combine it with domain expertise and derive insights. As an example, the 4000-page omnibus bill is "transparent" in principle and in a legal sense, but certainly not in a practical sense for most people. There's a lot more like it: laws, spending bills, federal budgets, freedom of information act responses, lobbying disclosures... Only a few highly trained professionals (investigative journalists) could historically process this information. This bottleneck might dissolve - not only are the professionals further empowered, but a lot more people can participate. Some examples to be precise: Detailed accounting of spending and budgets, diff tracking of legislation, individual voting trends w.r.t. stated positions or speeches, lobbying and influence (e.g. graph of lobbyist -> firm -> client -> legislator -> committee -> vote -> regulation), procurement and contracting, regulatory capture warning lights, judicial and legal patterns, campaign finance... Local governments might be even more interesting because the governed population is smaller so there is less national coverage: city council meetings, decisions around zoning, policing, schools, utilities... Certainly, the same tools can easily cut the other way and it's worth being very mindful of that, but I lean optimistic overall that added participation, transparency and accountability will improve democratic, free societies. (the quoted tweet is half-ish related, but inspired me to post some recent thoughts)

Gemma 4 26B-A4B (MoE) benchmark results are out, judge yourself! - APEX Balanced (18.1 GB) crushes on PPL (316.4 vs Q8_0's 337.3!) and beats all baselines on Winogrande (54.0%), MMLU (26.4), ARC (28.4) - APEX beats @UnslothAI quants on every accuracy metric at similar sizes - All APEX variants are 10-18% faster than baselines on tg128

This is the shot you can’t get from the press site. This camera was sitting a few football fields from the SLS rocket at Pad 39B for days before launch, baking in the Florida sun, surviving rain, humidity, and whatever else the Cape threw at it. No photographer behind the viewfinder. Just a camera, a sound trigger, and a bet. The way pad remotes work: you set your camera up days in advance, dial in your composition, lock everything down, and walk away. You don’t touch it again until after the launch. The shutter fires on sound activation with a @MiopsTrigger smart+ trigger. With SLS, the four RS-25 engines ignite six seconds before the solid rocket boosters, so the camera is already firing before the vehicle even leaves the pad. You get home, pull the card, and find out if you nailed it or if a bird landed on your lens two days ago and left your a present and you got 400 photos of soemthing crappy. There’s no formula for protecting your gear this close. Some photographers build wooden boxes with doors that pop open. Some use plastic bags and tape. Some do plastic or metal barn door rigs on hinges. I tend to leave mine open just in plastic rain covers because boxes limit my composition and setup time, but that means your cameras are more exposed to the elements and whatever energy and debris comes off the pad. You’re basically gambling a camera body every time you set one. That’s what I love about this genre. There’s no playbook. You make it up as you go. Every time is an adventure. 📸 credit: me for @SuperclusterHQ - Artemis II pad remote | ~1,000 ft from Pad 39B | Kennedy Space Center

Emprender con socios es una de las decisiones más importantes que tomarás en tu vida profesional. No puede faltar el pacto de socios. finanboo.com/blog/articulos…

Meet Gemma 4: our new family of open models you can run on your own hardware. Built for advanced reasoning and agentic workflows, we’re releasing them under an Apache 2.0 license. Here’s what’s new 🧵

Wrote a small post about this idea. davidgasquez.com/reliable-unrel…

We want to apply APEX to every MoE model out there! But we need GPU time to run the benchmarks (each model takes 6-8 hours of eval across all metrics). If you have spare GPU cycles and want to help, DM me. We'll credit contributors and share results publicly. Let's make every MoE model run on consumer hardware!

Qwen3.5-9B-OmniCoder-Claude-Polaris-Text-dwq4-mlx benchmarked on a M1 Ultra: MBPP 63.8% - 824.4s LIVECODEBENCH 26.7% - 26789s #LocalLLM

El otro día se me ocurrió un proyecto: Montar en un pendrive arrancable todo lo que necesites en caso de emergencia (apagón, incendio, catástrofe climática...), desde aplicaciones que corran 100% offline hasta tus datos personales. Así ha nacido refugiOS github.com/Ganso/refugiOS/