This is not the GMKtec EVO-X2 it’s the Ryzen AI Halo. Also, in no way does it replace a $200 per month 2T parameter model frontier cloud AI sub. All of this stuff has value but we have to try to be accurate about what that value is or people will buy it and say they got rugged. That’s not helping anybody.

Anthropic is evil and this is part of their plan x.com/TheAhmadOsman/status/2…

Congrats to @GoogleDeepMind on the launch of DiffusionGemma. The model generates 256 tokens in parallel per step, delivering 150 TPS on DGX Spark, and 1,000 TPS on a single H100. We're supporting it from day one with: • BF16 and NVFP4 checkpoints on @huggingface🤗 • Free GPU-accelerated endpoints on build.nvidia.com • @vllm_project support with FP8 precision Get started with DiffusionGemma on NVIDIA: nvda.ws/43ro19u

This year opensourceaimustwin.com/?sha… x.com/theahmadosman/status/2…

If you're running with 24GB VRAM or unified memory, this is for you! x.com/mr_r0b0t/status/206436…

Hello Everyone 👋 I’m building DreamOS, a minimal Linux-based operating system focused entirely on local AI inference. It boots directly into a dashboard for chatting, model management, benchmarks and live hardware metrics. The core control plane is written in Rust. DreamOS uses the official NVIDIA Linux/CUDA stack, while owning the layers above it: hardware-aware configuration, model residency, KV-cache management, inference scheduling, context handling, telemetry and automatic backend tuning. The current development target is Qwen3.5-9B Q4_K_M on an RTX 4060 with a 32K context. Initial runtime experiments have reached up to approximately 45 tokens/s, but I’m still building matched, reproducible benchmarks before claiming a definitive improvement over standard llama.cpp. I’m now profiling the exact CUDA decode bottlenecks, experimenting with custom Q4_K kernels, KV-cache compression, speculative decoding and persistent execution. There is also a longer-term bare-metal DreamOS research track, but the practical product comes first: making local models run faster, remain loaded, consume less memory and feel like a complete AI-native system rather than another application running on a general-purpose desktop OS. *Illustrative image

Se o Brasil de fato bloquear as IA's por causa de eleição, oque é péssimo, ao menos temos a alternativa de fazer as pessoas entenderem que estudar sobre Local AI é o futuro, cada vez mais modelos estão ficando melhores, mais rapidos e com menos necessidade de rodar em ''super computadores''

Geeking out about local AI and the mission of Light Heart Labs with @liamjf444 on his Podcast 💪❤️ youtu.be/LfnnwDQnQ-Y?si=vGyP…

Fun LLM Question from Mike today > Let’s say Thanos snaps away every model from existence except Qwen3.5-2B > Ahmad, what would you do? Assumptions & Clarifications 1. All papers, tech reports, synthetic datasets, HF repos, checkpoints, eval harnesses, and research artifacts are gone from existence 2. I still have my own knowledge and experience all my hardware 3. Inference Engine for Qwen3.5-2B is available to me What I would do… Phase I: - Immediately write down every training, inference, evaluation, data generation, and scaling trick I can remember before I forget details - Build a highly-deterministic synthetic data harness around Qwen3.5-2B - Generate purpose-specific datasets, starting with synthetic data generation - Finetune Qwen3.5-2B for generating better synthetic data Phase II: - Finetune for specialized versions on coding, research, agents, and kernels using further generated synthetic data from the finetuned model - Create automated research loops that continuously improve datasets and training runs - Scale what works and stop what doesn’t Phase III: - Train a 9B-class model using the above - Finetune a highly specialized 9B-class model that dominates a narrow domain, focusing on high-return $$$ markets Phase IV: - Raise capital to Buy massive amounts of GPUs - Hire researchers - Repeat with scaling up as the primary goal Kinda simplified plan but I thought this was an interesting one to share

We should switch to certification and licensing systems for professions. You go through standardized and recognized tests of competency in a field or area, and if you have the skills and abilities and knowledge it should be wholly irrelevant how you learned it all and got there.

Colleges and universities are over priced and often unnecessary middlemen for what real learning and growth looks like in 2026.

It was good enough for me to clear up some doubts before adapting it to my own kernel.

Finally, today's Nemotron 3 Ultra release makes me very hopeful for the future of Opensource AI Jensen knows that this is important to keep the powers in check, and I believe he's sincere in his answer to me that there will be continuity to the Nemotron Coalition releases Big W

This is what progress looks like, friends. 😎 Let's hope we get better support for our RTX Blackwell cards and DGX Sparks soon!

🚨Jensen Huang gifted Faker a one-of-a-kind graphics card personally signed by him. “Only one in the world. This might be worth a million dollars. I might have to keep this now.” The king of AI handing a legendary gift to the king of League. A truly iconic moment.

Meet Gemma 4 12B! A unified, encoder-free multimodal model designed to bring high-performance intelligence directly to your laptop, and released under an Apache 2.0 license. Bridging the gap between edge efficiency and advanced reasoning. Here is what’s new with Gemma 4 12B: 👇

14x RTX 3090s Qwen 3.6 27B Running 42 agents IN PARALLEL at full 256k context - exl3 6bpw - fp8 KV Cache - Aphrodite Inference Engine w/ tp=2, pp=7 The world of agents will run locally btw

Neymar via instagram. The last dance. 🙏🏾🇧🇷 Vamos buscar

Introducing Cosmos 3: Our latest frontier model for Physical AI Cosmos 3 is the world’s first fully open omnimodel with native vision reasoning, world and action generation. Today we’re releasing Super (32B) and Nano (8B) variants.

Entire world: We need more GPUs Meanwhile, Jensen Huang: