The retro handheld market has a strange problem. The hardware keeps getting better, the screens get sharper, the processors get faster, and yet most of these devices land looking like prototypes someone forgot to finish. Generic shells, forgettable proportions, and LED lighting as a substitute for actual design thinking. For a category built entirely on nostalgia, very few of these devices actually look like they belong to any era at all.

That tension is what one Reddit user decided to address. Starting with a Retroid Pocket 5, a $199 Android handheld running a Snapdragon 865 and a 5.5-inch AMOLED display, the mod layers Sony and Nintendo branding onto the same shell. Vinyl decals, translucent polycarbonate, a 3D-printed volume rocker from Etsy, and a cable replaced in PS2 color. The result looks less like a sticker job and more like a concept render from an alternate 1999.

Designer: Mitchieyan

The translucent shell is doing most of the work. It pulls from the visual language of the N64’s Funtastic series, those clear and atomic-purple controllers Nintendo released in the late 1990s, where showing the circuitry was the design choice rather than concealing it. Over a piano-black grip body with PlayStation-colored face buttons, the frosted polycarbonate shifts from grey to near-white depending on the light. It shouldn’t feel considered. It does.

The branding placement is where intent becomes clear. The Sony wordmark sits centered on the upper face, exactly where it appeared on a PSOne. Below it, the PlayStation four-color logo. At the bottom bezel, the Nintendo badge mirrors its position on a Game Boy Advance SP. None of it is licensed, of course. These are adhesive vinyls placed by someone who grew up with both systems and wanted their coexistence on one device to feel inevitable rather than absurd.

Not everything here reaches backward. The analog sticks are translucent caps over hall-effect sensors, lit teal on the left and purple on the right, owing nothing to 1999. That generation didn’t have RGB anything. The lighting reads as a concession to the present; the one feature announcing this is still an Android device in 2025, not a prototype from some alternate Sony-Nintendo licensing meeting. Whether it sits comfortably alongside the retro shell is a fair question.

The rear view shifts the frame again. A large dual-grip body in smooth black rubber dominates the back, a clear plastic hinge connecting the screen to grip in full view, structural and unapologetic. The 3D-printed volume rocker at the top edge puts a physical control where fingers naturally land. The back half feels closer to a DualShock than a Game Boy, which is either the point or the problem, depending on what you wanted this thing to be.

Flip to the front screen, and the emulator grid makes the whole thing literal. DuckStation for PS1, Dolphin for GameCube, PPSSPP for PSP, melonDS for Nintendo DS, and a live PS2 wallpaper cycling behind all of it. This device runs both companies’ libraries simultaneously without asking permission from either. The branding on the shell, in that context, stops being a novelty and starts reading as a plain statement of what the hardware already does.

The retro handheld category is large enough now that sameness has become its default. The Retroid Pocket 6, the current flagship from the same manufacturer, drew community criticism for being indistinguishable from competitors: glass front, LED sticks, rounded edges, and no particular character. A fan mod building identity out of borrowed logos is one response to a problem the manufacturers haven’t solved. It’s also just someone enjoying a hobby and being honest about what they want.

The hardware to play PS1, PS2, GameCube, and Game Boy Advance all on one screen already exists and costs under $200. What the market hasn’t resolved is what that device should actually look like, or whose name should go on it. This mod doesn’t answer either question. It just makes the gap between what’s technically possible and what anyone has bothered to design feel a little harder to dismiss.

The post This Fan Made the Sony-Nintendo Handheld the Companies Never Would first appeared on Yanko Design.

At GTC 2026, Jensen Huang, Aravind Srinivas, Harrison Chase, Mira Murati, and Michael Truell made a compelling case that the future of AI belongs to open agent systems, not just open models.

The post Nvidia’s Open Model Super Panel Made a Strong Case for Open Agents appeared first on TechRepublic.

Learn more about how AI agents are reshaping enterprise workflows in TechRepublic’s latest generative AI coverage.

The post Xiaomi Unmasks Mystery AI Model Behind DeepSeek Buzz appeared first on TechRepublic.

llamafile est un projet complètement barré qui va vous permettre de transformer des modèles de langage en exécutables. Derrière se cache en fait la fusion de deux projets bien badass : llama.cpp , un framework open source de chatbot IA, et Cosmopolitan Libc , une libc portable pour compiler des programmes C multiplateformes. En combinant astucieusement ces deux technos, les petits gars de Mozilla ont réussi à pondre un outil qui transforme les poids de modèles de langage naturel en binaires exécutables.

Imaginez un peu, vous avez un modèle de langage qui pèse dans les 4 gigas, dans un format .gguf (un format couramment utilisé pour les poids de LLM). Et bien avec llamafile, vous pouvez le transformer en un exécutable standalone qui fonctionnera directement sur le système sur lequel il est sans avoir besoin d'installer quoi que ce soit. Ça va permettre de démocratiser l'utilisation et la diffusion des LLM.

Et niveau portabilité, c'est le feu puisque ça tourne sur six OS, de Windows à FreeBSD en passant par macOS. Les devs ont bien bossé pour que ça passe partout, en résolvant des trucs bien crados comme le support des GPU et de dlopen() dans Cosmopolitan et croyez-moi (enfin, croyez-les) ça n'a pas été une mince affaire !

Niveau perf aussi c'est du brutal ! Sur Linux llamafile utilise pledge() et SECCOMP pour sandboxer le bousin et empêcher les accès fichiers non désirés et avec les derniers patchs de Justine Tunney , la perf CPU pour l'inférence en local a pris un boost de malade du genre 10 fois plus rapide qu'avant. Même sur un Raspberry Pi on peut faire tourner des petits modèles à une vitesse honnête.

Mise à jour : llamafile 0.10

Bonne nouvelle, le projet est loin d'être mort puisque la version 0.10 vient de sortir (mars 2026) et elle apporte pas mal de changements. Déjà, le projet a migré de Mozilla Ocho vers Mozilla.ai , ce qui montre que Mozilla prend le truc au sérieux côté IA.

Le gros morceau de cette release, c'est un tout nouveau build system. Fini le bazar monolithique, maintenant llama.cpp, whisper.cpp et Stable Diffusion sont intégrés comme des sous-modules Git. L'avantage c'est que ça permet de suivre beaucoup plus facilement les dernières versions de llama.cpp et donc de supporter les modèles les plus récents dès leur sortie.

Côté utilisation, on a maintenant trois modes bien distincts :

• Mode TUI (Terminal User Interface) : vous chattez directement dans votre terminal avec le modèle, avec même un mode "think" pour le raisonnement étendu
• Mode CLI : pour poser une question rapide en one-shot, genre llamafile "c'est quoi un llamafile ?" et hop, la réponse arrive direct
• Mode serveur : avec le flag --server, ça lance le serveur llama.cpp classique pour exposer une API compatible OpenAI

Autre truc cool, le support multimodal est là avec le nouvel argument --image. Vous pouvez balancer une image au modèle et il l'analyse. Ça marche avec des modèles comme Qwen3-VL, LLaVA 1.6 ou Ministral 3.

Côté GPU, Metal fonctionne nativement sur macOS (ARM64) sans bidouille, et le support CUDA est restauré sur Linux. Par contre, le GPU sur Windows n'est pas encore de la partie, et le sandboxing via pledge()/SECCOMP a été temporairement retiré dans cette version.

Bref, si vous aviez testé llamafile il y a un moment et que vous aviez trouvé ça un peu limité, c'est peut-être le moment de retélécharger la bête et de voir ce que ça donne avec les modèles de 2026. C'est toujours aussi simple : un fichier, on le rend exécutable, on le lance, et c'est parti.

Alors on dit merci qui ?

Merci Mozilla ! 🙏🦊

Amis crafteurs, préparez-vous à voir votre monde d'une toute nouvelle façon car il est désormais possible de recréer votre ville, votre quartier ou même la Cathédrale de Clermont-Ferrand dans Minecraft (Java ET Bedrock) avec une précision chirurgicale. C'est ce que permet de faire Arnis , un projet open source vraiment très cool.

Développé en Rust, Arnis fait le pont entre notre bonne vieille Terre et l'univers cubique de Minecraft en exploitant tout simplement la puissance d'OpenStreetMap. Pour rappel, c'est une formidable base de données cartographique collaborative qui recense routes, bâtiments et points d'intérêt du monde entier.

Le principe est donc simple mais efficace : vous sélectionnez une zone géographique avec l'outil de sélection intégré, vous choisissez votre monde Minecraft (de préférence un monde plat), et Arnis se chargera automatiquement de :

1. Récupérer les données géographiques via l'API Overpass
2. Convertir ces informations en coordonnées Minecraft
3. Traiter les éléments par ordre de priorité pour éviter les conflits
4. Générer une couche de sol adaptée avec les données d'élévation
5. Construire les structures bloc par bloc
6. Sauvegarder le tout dans votre monde

La beauté d'Arnis réside dans sa capacité à retranscrire les moindres détails. Les bâtiments conservent leurs proportions, les routes suivent leur tracé réel, et même les zones vertes sont fidèlement reproduites. Le système gère intelligemment les bâtiments avec leur hauteur relative, les routes et autoroutes, les espaces verts et plans d'eau, les points d'intérêt particuliers et tout ce qui est fontaines et structures spéciales.

Niveau personnalisation, Arnis propose également :

• Un mode Hiver pour transformer votre zone en paysage enneigé
• Un facteur d'échelle pour ajuster la taille de votre génération
• Un réglage de la hauteur du sol pour définir l'altitude de base
• Un Timeout Floodfill pour contrôler le temps de génération des zones d'eau
• Et une BBOX personnalisée permettant d'entrer manuellement les coordonnées de la zone désirée

D'ailleurs, la version 2.5 sortie en février 2026 (baptisée "Metropolis Update") a apporté pas mal de nouveautés bien senties. Déjà, Arnis supporte maintenant Minecraft Bedrock en plus de Java, du coup ça marche aussi sur console et mobile. Ensuite, le générateur intègre les données d'élévation pour un terrain réaliste avec de vraies collines et vallées... plus besoin de se contenter d'un monde plat. Et le truc franchement cool, c'est la génération d'intérieurs de bâtiments avec du mobilier. Pour les grosses zones ou si vous êtes sur mobile, y'a aussi MapSmith, une alternative qui tourne directement dans le navigateur.

Pour tirer le meilleur parti d'Arnis, vous devrez utiliser Minecraft 1.17 minimum (Java) ou la dernière version Bedrock, initialiser un monde tout plat pour éviter les conflits de terrain, et une fois la génération terminée, vous n'aurez plus qu'à vous téléporter aux coordonnées 0 0 0 (/tp 0 0 0). Pensez à explorer vers les X et Z positifs si vous ne voyez rien immédiatement et si vous êtes sur Windows, installez l'Evergreen Bootstrapper de Microsoft si nécessaire.

La communauté bosse activement sur le projet. Et si vous voulez pas installer Rust et compiler, y'a des binaires prêts à l'emploi sur la page des releases . Hop, on télécharge, on lance, et on se balade dans son quartier version cubique !

Article publié initialement le 11 février 2025 et mis à jour le 19 mars 2026.

Most people who game on a PC own two things that do roughly the same job at different times: a mouse for the desk and a gamepad for the couch. They live side by side, occasionally getting in each other’s way, and neither one is going anywhere. Pixelpaw Labs, a hardware startup from Bangalore, India, thinks that arrangement is wasteful and has built something to prove it.

The Phase is a wireless mouse that physically separates down the middle into two independent halves. Snapped together, it sits on a desk and works like a normal mouse. Pull it apart, and each half reveals a joystick, triggers, a D-pad on the left side, and face buttons on the right, a split gamepad that was hiding in plain sight the whole time.

Designer: Pixelpaw Labs

That missing scroll wheel is not an oversight. Fitting a traditional wheel in the center of the body would have made the split mechanism impossible, so Pixelpaw replaced it with a capacitive touch strip along the top of the left button. Flicking a finger across it scrolls through documents and web pages, with a glide feature that lets the momentum coast rather than stop abruptly. It’s a trade-off that works around a real geometric constraint.

As a mouse, the Phase is competitive on paper. A 16,000 DPI optical sensor pairs with a 1,000 Hz polling rate when connected via the included 2.4 GHz USB dongle. Bluetooth LE is available for convenience and multi-device pairing across up to three devices, though the polling rate drops to 125 Hz in that mode, a gap that matters in fast-paced PC games.

Up to 18 customizable buttons are mappable through the Pixelplay companion app, and a Layer button doubles each button’s function capacity without adding physical complexity. Battery life is rated at 72 hours per charge over USB-C, which is more than enough to outlast dedicated gaming sessions on either side of its personality.

The controller halves use mechanical tactile switches, which is more than most mobile gaming clip-ons bother with. Pixelpaw also has an accessory called the Phasegrip, a bracket that holds the two separated halves apart with a smartphone mounted in the center, turning the setup into a handheld console for mobile gaming. The Phase works across PC, Android, iOS, iPadOS, and ChromeOS, so switching between devices doesn’t require swapping hardware.

Everything shown so far is pre-production, and the company has been upfront that the final surface finish will differ. That’s a meaningful caveat for a product whose physical fit and feel will determine whether the concept actually holds up. Whether they’ll be able to deliver this Holy Grail of PC gaming, however, is the real question that can only be answered in time.

The post This Wireless Mouse Splits in Half to Reveal a Hidden Game Controller first appeared on Yanko Design.

Most garden structures ask one thing of you: sit still and enjoy the shade. A pergola is a pergola, a gazebo is a gazebo, and neither one particularly cares what the afternoon light is doing. Michael Jantzen’s Interactive Garden Pavilion operates on a different premise entirely, one where the occupant has as much say over the structure as the designer did.

Built from sustainably grown stained wood and painted a uniform forest green, the pavilion sits on an octagonal support frame fitted with 30 slatted hinged panels across its walls and roof. Each panel pivots independently, sliding and rotating along the frame before locking into position. Open them wide on a hot afternoon, and the interior breathes. Angle them down against the glare, and the space dims considerably.

Designer: Jantzen

That last point is where the design earns its name. Most adjustable outdoor structures offer a single variable, usually an awning or a retractable canopy, within an otherwise fixed form. Here, the entire skin of the building is the variable. The wall panels, roof panels, and ground-level platform extensions can all be repositioned, which means the pavilion can look substantially different from one afternoon to the next.

Pull the panels shut on three sides, and the structure becomes a genuinely private enclosure. Splay them open, and the interior connects fully to the garden around it. In one arrangement, it reads as a dense closed form. In another, the structure opens up entirely, and the slatted framework becomes almost sculptural against the lawn.

Inside, two benches with adjustable backrests run the length of the interior, facing each other. The seating is built into the frame, which keeps the floor plan clean and leaves room to recline fully. When the overhead panels are partially open, sunlight enters in sharp parallel bands that shift across the benches as the day moves, a quality that is either meditative or distracting depending on what you came in for.

The construction logic is also notably practical. The pavilion is a prefabricated modular system, so the components can be scaled before assembly or joined with additional units to form a larger cluster. No foundation is required in most configurations. Given its size and type, a building permit is unlikely to be needed in many jurisdictions, which removes one of the more tedious barriers between an interesting design and an actual garden.

Jantzen has spent decades proposing architecture that responds dynamically to its occupants, much of it remaining on paper. This pavilion is one of the cases where the idea got built, and the result holds up at close range. The slatted wood is honest about what it is, the green paint ties the structure to the garden without trying to disappear into it, and the hinge mechanism does exactly what it promises.

The post Michael Jantzen’s Garden Retreat Has 30 Panels to Rearrange by Hand first appeared on Yanko Design.

How a small IT team at the South Mississippi Housing Authority used AI and hybrid cloud technology to handle 68% of calls and modernize legacy infrastructure.

The post What IT Leaders Can Learn From a Housing Authority’s AI Transformation appeared first on TechRepublic.

Microsoft says Xbox mode is coming to Windows 11 in April, bringing a console-style interface, faster game loading tools, and a unified PC strategy.

The post Microsoft Plans ‘Xbox Mode’ for Windows 11 PCs Starting in April appeared first on TechRepublic.

Google vient de lancer Gemini Embedding 2, son premier modèle d'embedding nativement multimodal. Texte, images, vidéo, audio et documents sont projetés dans un même espace vectoriel, ce qui permet de faire de la recherche sémantique croisée entre différents types de contenus.

Un seul modèle pour tout indexer

Jusqu'à présent, les modèles d'embedding se limitaient au texte. Vous vouliez indexer des images ou de la vidéo, il fallait un autre pipeline. Gemini Embedding 2 fait tout d'un coup : vous lui envoyez du texte, des images (jusqu'à 6), de la vidéo (jusqu'à 120 secondes) ou de l'audio (jusqu'à 80 secondes), et il vous renvoie un vecteur dans le même espace. Le modèle gère plus de 100 langues et prend en charge jusqu'à 8 192 tokens en entrée pour le texte.

Côté technique, le modèle utilise le Matryoshka Representation Learning, ce qui permet de choisir la taille des embeddings entre 128 et 3 072 dimensions. Google recommande 768 dimensions pour un bon compromis entre qualité et stockage, ce qui divise par quatre l'espace disque par rapport à la taille maximale.

Les tarifs et la concurrence

Le texte est facturé 0,20 dollar par million de tokens, avec un mode batch à moitié prix. Les images montent à 0,45 dollar, l'audio à 6,50 dollars et la vidéo à 12 dollars par million de tokens. Un palier gratuit est disponible pour tester.

Côté performances, Google affiche de bons scores sur les benchmarks MTEB : 69,9 en multilingue et 84,0 en code. Mais pour du texte seul, OpenAI reste bien moins cher avec son text-embedding-3-small à 0,02 dollar par million de tokens, soit dix fois moins.

Le modèle est disponible via l'API Gemini et Vertex AI, et compatible avec LangChain, LlamaIndex, Weaviate ou ChromaDB.

Le vrai argument de Google ici, c'est le multimodal. Si vous avez besoin d'indexer des catalogues produits avec photos et descriptions dans le même vecteur, ou de faire de la recherche dans des archives vidéo, il n'y a pas d'équivalent chez OpenAI pour le moment.

Mais pour du texte pur, la différence de prix est quand même importante. On attend de voir comment ça se comporte en production, et si les scores MTEB se confirment sur des cas d'usage réels.

Source : Blog Google

The combination of rare earth materials and cryogenic temperatures could be the key to transforming power infrastructures in preparing for AI applications.

The post High-Temperature Superconductors Could Redefine Data Center Power Density appeared first on TechRepublic.

Anthropic rolls out Claude Sonnet 4.6 as its new default model, bringing stronger reasoning and coding power to free and paid users alike.

The post Anthropic Launches Claude Sonnet 4.6 as Default Model for Free and Paid Users appeared first on TechRepublic.

What if we designed homes the way cats would design them? Not human homes with a token scratching post in the corner but true spatial systems built around curiosity, vertical exploration, territorial comfort, and play. The N Plus Magic House begins precisely at that question, reframing pet furniture not as an accessory but as architecture scaled for feline psychology. Instead of treating a cat house as a static object, this project treats it as a living spatial framework, one that evolves alongside its inhabitant.

Today’s pet owners increasingly see their cats as emotional companions rather than animals that merely coexist in domestic space. That shift has quietly created a design problem. Traditional cat houses, even elaborate ones, tend to be fixed structures. They may be visually impressive, but they impose constraints on placement, adaptability, and long-term usability. The N Plus Magic House flips that paradigm by introducing modularity as its core philosophy. Rather than selling a finished form, it offers a system of standardized units that can be assembled, rearranged, expanded, or reduced as needed. The result is less like furniture and more like a customizable habitat kit.

Designer: Taizhou Hake Technology Co., Ltd

The genius of the design lies in its simplicity. Each module functions independently yet connects securely through precision-engineered connectors. Owners assemble structures by inserting panels into slots and stacking them like building blocks. No technical expertise, tools, or installation manuals are required. This intuitive construction method does something subtle but powerful. It turns pet care into participation. Instead of buying a finished object, users become co-designers of their cat’s environment. That interaction strengthens the emotional bond among the owner, the pet, and the space.

Material choices reinforce the system’s practicality. The structure combines impact-resistant PP resin, transparent PET panels for visibility, and carbon steel mesh for structural integrity. These materials balance durability with safety while allowing owners to monitor their pets without disturbing them. The manufacturing processes, such as injection molding and automatic wire welding, ensure consistency, precision, and reliability across units. Every element reflects careful alignment with feline behavior and safety requirements.

Behind the scenes, the development team approached the project with a research-driven mindset. They studied cats’ behavioral patterns, analyzed existing products on the market, and mapped owner expectations. One of the biggest technical challenges was maintaining structural stability while preserving modular flexibility. The solution was a custom connector engineered to withstand pressure and weight while preventing slippage. Its textured surface increases friction, ensuring modules remain firmly locked during use. This small component is arguably the system’s unsung hero. It transforms a playful concept into a reliable architectural structure.

Developed in Taizhou, Zhejiang Province, between July 2023 and November 2024 and later exhibited internationally, the N Plus Magic House represents a broader shift in product design thinking. It signals a move away from static ownership toward adaptive systems, objects that respond to changing needs over time. In a world where personalization defines modern consumer expectations, this approach feels less like a novelty and more like a glimpse into the future of domestic product design.

The post The Modular Cat Habitat That Turns Playful Curiosity Into Living Architecture first appeared on Yanko Design.

Screens are on all day, and hands tend to find something to do when focus slips. Pen clicking, phone picking, knuckle cracking, the nervous tics of modern desk work. Most “mindfulness” solutions are still apps, which is a bit ironic when the problem is too much screen time. There’s something to be said for a small mechanical object that gives your brain a reset without asking for any attention in return.

Amsterdam Dynamics’ ST-01 is a modular spinning top and tactile focus object built for desks, hands, and minds that rarely get a break. It’s intentionally simple but not single-purpose, offering multiple mechanical interactions with no correct sequence. You use one when you need it or work through all of them. No app, no setup, no instructions, just the object and whatever your hands feel like doing with it.

Designer: Antonio Lo Presti (Amsterdam Dynamics)

A gentle twist sets ST-01 rotating smoothly, and it’s engineered to keep going for over two minutes. That’s long enough to watch while thinking through a problem, waiting for a file to render, or cooling down after a difficult meeting. It functions as a visual anchor, something calm and physical in a field of notifications and browser tabs that all want something from you.

Pressing the top triggers a satisfying mechanical click with clear tactile feedback, the kind of repeatable, purposeful sensation that replaces the nervous habit of clicking pens or tapping a keyboard. Amsterdam Dynamics calls it “reset focus,” which is accurate if not exactly humble. The middle disc can also be flipped like a coin, adding a small decision-making tool and another texture to the interaction when you need a nudge in either direction.

Of course, there’s a modular construction underneath all of that. ST-01 is built from three parts that can be taken apart and snapped back together, held in alignment by two precisely positioned magnets. That magnetic core keeps the structure stable during spinning while making it easy to disassemble by hand. There’s also a built-in magnet that lets it stick to metal surfaces, which is either a neat trick or a genuinely useful parking spot depending on your desk.

CNC-machined, anodized aerospace-grade aluminum means it’s solid in the hand, balanced, and finished in a way you notice the first time you pick it up. Cheap fidget toys flex, squeak, and wear out quickly. ST-01 is designed to stay on the desk for a long time, with Amsterdam Dynamics framing it as “a beautiful object, made to last a lifetime” and something that can eventually be passed on.

That’s an unusual positioning for what is essentially a desk toy, but it fits the overall idea. ST-01 doesn’t ask for a lifestyle change, a daily streak, or a subscription. It just gives your hands a few repeatable interactions and a place to return when the work gets loud, which turns out to be the kind of quiet, mechanical company a desk actually needs.

The post Forget Mindfulness Apps, This Desk Top Spins for 2 Minutes Instead first appeared on Yanko Design.

UpScrolled says it has passed 2.5 million users after TikTok’s US ownership change, as the rival app scrambles to scale and set moderation rules.

The post This New Social Network Says It Won’t Sell Your Data — and It Just Hit 2.5M Users appeared first on TechRepublic.

Remember when buying a sofa meant measuring doorways, hoping it would fit in the elevator, and bribing friends with pizza to help you haul it up three flights of stairs? Designer Yuqi Wang just made all of that obsolete with the Accordion Modular Sofa, and the concept is as brilliant as it sounds.

The whole design centers on one genius idea: what if you only needed one type of module to create any sofa configuration you could imagine? The Accordion does exactly that with a single base unit that compresses and expands like its musical namesake. Through an internal structure made from high-resilience foam, wooden side panels, climbing ropes, and adjustable knobs, each module can change its size and curvature to fit whatever space you’re working with.

Designer: Yuqi Wang

What makes this really clever is how it solves multiple problems at once. Traditional modular sofas usually require you to buy different shaped pieces (corner units, armrests, middle sections) and hope you’ve guessed right about what you need. With the Accordion system, you’re working with identical modules that transform as needed. Want a compact loveseat for your apartment? Use two modules. Need a sprawling sectional for your living room? Add more units and configure them however you like. The possibilities aren’t just numerous, they’re practically endless.

But here’s where it gets even better. Because these modules compress, the whole thing ships flat-packed in a way that actually makes sense. We’re talking small enough to literally place in a shopping cart and carry out yourself. No more waiting weeks for delivery windows or paying outrageous shipping fees. The compression feature dramatically cuts down on packaging waste, storage costs, and the carbon footprint of transportation. It’s the kind of sustainable design choice that doesn’t require you to sacrifice anything in return.

The inspiration came from watching an accordion player perform, where Wang noticed how the instrument’s bellows compress and expand with such fluidity. That rhythmic movement translated into furniture creates something that feels almost alive, like the sofa adapts to you rather than the other way around. The technical execution involved studying compression mechanisms from various industrial applications before landing on a prototype that was both simple and effective.

For anyone who’s moved apartments multiple times or likes to rearrange furniture seasonally, this is a game changer. Your sofa can literally grow or shrink with your needs. Hosting a party? Expand it. Need more floor space for a home workout? Compress it down. The adjustable knobs make reconfiguring straightforward enough that you won’t need tools or an engineering degree.

The design world has taken notice. The Accordion Modular Sofa won the Golden A’ Design Award in Furniture Design for 2025, one of the most prestigious recognitions in the field. This isn’t just a participation trophy either. The Golden award represents first-rate, outstanding design that genuinely advances the intersection of art, science, and technology. What’s exciting is that this isn’t some far-off concept. The sofa is scheduled to hit the market in October 2025, which means you could actually have one of these in your home soon. It represents a shift in how we think about furniture ownership, moving away from bulky investments that lock you into one configuration toward adaptable pieces that evolve with your lifestyle.

The Accordion Modular Sofa proves that innovative design doesn’t have to be complicated or precious. Sometimes the best ideas are the ones that make you wonder why nobody thought of them sooner. A sofa that compresses for transport, expands for comfort, and reconfigures endlessly from a single module type? That’s the kind of practical magic that makes design truly great.

The post This New Sofa Compresses Like an Accordion (Fits in Your Cart) first appeared on Yanko Design.

Vous voulez faire tourner un modèle d'IA en local sans avoir besoin d'un serveur de la NASA ? Eh bien Nvidia vient de lâcher une bombe avec Nemotron 3, une famille de modèles open source plutôt impressionnant et surtout, ils ont publié leurs données d'entraînement afin de jouer la transparence totale. Chapeau !

Le modèle phare de cette nouvelle famille s'appelle Nemotron 3 Nano et c'est un modèle de 30 milliards de paramètres, mais attention, il n'en active que 3,5 milliards à la fois grâce à une architecture hybride qui mélange du Mamba-2 et du Mixture-of-Experts ( MoE ). Ça permet de garder des performances de ouf tout en restant léger niveau ressources.

Sous le capot, Nvidia a également mis le paquet puisque le modèle a été entraîné sur 25 trillions de tokens. J'ai bien dit "trillions"... Pour vous donner une idée, les données d'entraînement incluent du Common Crawl de 2013 à 2025, du code dans 43 langages différents, des articles scientifiques, et une tonne de données synthétiques générées par d'autres modèles. Et tout ça, Nvidia l'a rendu public donc vous pouvez télécharger les datasets sur Hugging Face et vérifier par vous-même ce qui a servi à entraîner le bouzin.

Côté performances, Nemotron 3 Nano se défend plutôt bien . Sur les benchmarks de raisonnement mathématique comme AIME25, il atteint 99,2% quand on lui donne accès à des outils. Sur le coding avec LiveCodeBench, il tape du 68,3%, ce qui le place devant Qwen3-30B. Et pour les tâches d'agent logiciel genre SWE-Bench, il monte à 38,8%. Pas mal pour un modèle qu'on peut faire tourner sur du matos grand public.

D'ailleurs, parlons du matos justement. Nemotron 3 Nano tourne sur des cartes comme la H100, la A100, ou même la future RTX PRO 6000 et supporte jusqu'à 1 million de tokens en contexte si vous avez assez de VRAM. Et niveau vitesse, Nvidia annonce un débit de tokens 4 fois supérieur à la génération précédente, avec 60% de tokens de raisonnement en moins. C'est donc exactement ce que tout le monde demande à saoir du token qui sort vite pour les workflows agentiques.

Maintenant, pour l'utiliser, c'est hyper simple. Il est dispo sur Hugging Face, et vous pouvez le lancer avec Transformers, vLLM, TensorRT, ou même llama.cpp. Y'a même un mode "thinking" qu'on peut activer ou désactiver selon si on veut du raisonnement poussé ou des réponses rapides.

Pour ma part, je l'ai testé à l'aide d'Ollama comme ceci :

ollama run nemotron-3-nano:30b

Offices keep buying furniture that looks permanent, which works fine until someone needs the room to do something different. A workshop space becomes a presentation area, a meeting room needs to turn into individual work zones, and nobody wants to wait three days for facilities to show up with screwdrivers. The furniture just sits there looking expensive and immovable while everyone works around it instead of with it.

PIXEL by Bene is designer Didi Lenz’s answer, and it looks almost suspiciously simple. Each piece is a 36 x 36 cm cube made from raw pine plywood with visible grain and knots all over the surface. Lenz says it isn’t really furniture, which makes sense when you see people stacking them into benches, flipping them into tables, or just using one as a side storage box with a handle cut into the side.

Designer: Didi Lenze (Bene)

The wood is completely untreated, so every cube looks slightly different depending on which part of the tree it came from. Some have dark knots near the corners, others show lighter grain patterns, and the plywood edges are exposed instead of hidden under veneer. It definitely reads as workshop material rather than corporate office product, which seems to be the whole point. You can see the screws holding the corners together.

The cubes stack easily because they’re all the same size, and the cutout handles on two sides let you carry them around or fold them over to connect boxes side by side. Add a white laminate top and a stack becomes a work table. Add casters to the bottom, and it rolls wherever you need it. PIXEL Rack adds metal frames that turn stacks into proper shelving or room dividers with slots for whiteboards and plants.

Bene shows photos of teams building entire project rooms by hand. Boxes stacked three high become benches for workshops, racks filled with boxes create semi-transparent walls between work zones, and tops laid across stacks turn into standing height tables. The setups look intentionally unfinished, like someone is still building them, which is probably the aesthetic Lenz wanted. Nothing looks bolted down or precious.

The system works because it assumes people will move things around themselves without asking permission. You need more seating for a presentation, so you grab some boxes from the storage wall and stack them into rows. The presentation ends, and those same boxes become side tables or go back to holding supplies. Heck, they can turn into a bar for an event if you add the right tops.

Raw plywood has obvious trade-offs. It’ll get dinged and stained over time, the surface isn’t smooth enough for detailed work, and the workshop look won’t suit every office brand. The fixed 36 cm dimension means everything is the same height whether you’re sitting, standing, or storing things, which can feel awkward. Some people will look at PIXEL and just see fancy storage crates, which isn’t entirely wrong.

But the system makes sense for spaces that need to change shape constantly. Co-working areas, design studios, classrooms, and pop-up shops can rebuild their layout between sessions without calling anyone. The wood looks honest and approachable instead of intimidating, and you don’t need instructions to figure out that boxes stack. PIXEL by Bene basically gives you building blocks that happen to be office furniture, or maybe it’s the other way around.

The post Bene Just Built Office Furniture You Can Reconfigure Without Any Tools first appeared on Yanko Design.

L’expert en sécurité Lander, a passé des semaines à cracker le format .lin de Splinter Cell sur la Xbox originale. C’est un format de fichier qui est resté mystérieux pendant deux décennies et c’est pas parce qu’il était ultra-sécurisé ou chiffré de fou.

Non, c’est juste parce qu’il était optimisé pour du matos obsolète !

En effet, les fichiers .lin, c’est le format utilisé par la version Xbox de Splinter Cell pour stocker tous les assets du jeu. Ce sont des archives compressées en zlib contenant des packages Unreal Engine 2, sauf que contrairement aux formats classiques que vous pouvez ouvrir tranquille avec n’importe quel outil d’extraction, ceux-là résistaient à tout depuis +20 ans. Hé oui, durant toutes ces années, la communauté modding essayait de les décoder mais sans succès…

On pensait que le problème venait du chiffrement mais en fait, pas du tout ! Le problème c’est que ce format était juste conçu d’une manière totalement inadaptée aux outils modernes. C’est ce que Lander appelle dans son article un “non-seekable format with lazy loading and interleaved exports”. En gros, vous pouviez pas juste pointer sur un offset et lire le fichier car tout était entrelacé, chargé à la demande, et optimisé pour la RAM limitée de la Xbox originale.

Du coup, impossible d’utiliser les extracteurs classiques d’archives UE2. La seule solution qu’il restait c’était donc de comprendre exactement comment le jeu chargeait ces fichiers en mémoire sur la console d’origine.

Lander a alors sorti l’artillerie lourde. D’abord l’émulateur xemu pour faire tourner la Xbox sur PC et ensuite il a posé des breakpoints mémoire pour capturer le moment exact où le jeu charge un fichier .lin.

Il a ensuite sorti IDA Pro pour analyser le code assemblé du jeu et quand l’analyse statique suffisait plus, il a patché directement le binaire du jeu pour forcer certains comportements…

Et bien sûr, pour notre plus grand plaisir, il documente tout ça dans un article avec des captures d’écran d’IDA Pro et des explications sur les structures de données. Bref, c’est du bon gros reverse engineering pur et dur, à l’ancienne, avec hex editor et la patience d’un moine bouddhiste.

Lander a ainsi réussi à extraire une partie du contenu tels que les textures, certains fichiers système, des données sur les niveaux du jeu, mais surtout, il a documenté la structure complète du format. Et ça, c’est peut-être plus important que l’extraction elle-même, parce que dans 10 ans, quand un autre passionné voudra bosser sur la préservation de Splinter Cell Xbox, il aura la doc. Il comprendra pourquoi ce format était comme ça, comment ça fonctionnait, quelles étaient les contraintes hardware de l’époque et j’en passe…

Voilà, une fois encore la préservation du jeu vidéo, c’est pas juste télécharger des ROMs mais c’est aussi comprendre les architectures, documenter les formats propriétaires, et reverse engineerer le code avant que ce ne soit trop tard !

Bravo Lander !