Image created with gemini-2.5-flash-image with claude-sonnet-4-5. Image prompt: A pristine 1961 Ferrari 250 GT California Spyder in Rosso Corsa red parked on a quiet tree-lined residential street at golden hour, warm porch lights glowing softly in the background, polished chrome reflecting neighborhood details, cinematic automotive photography with shallow depth of field, cozy community atmosphere with mailboxes and flowering window boxes visible, premium studio lighting with soft shadows and warm highlights, elegant composition emphasizing intimate local scale.
Claude can now create and use files \ Anthropic https://www.anthropic.com/news/create-files
🤯 400 Token/S on a MacBook? Yes, you read that right! Shaohong Chen just fine-tuned the Qwen3-0.6B LLM in under 2 minutes using Apple’s MLX framework. This is how you turn your MacBook into a serious LLM development rig. A step-by-step guide and performance metrics inside! 🧵 https://x.com/ModelScope2022/status/1977706364563865805
Meta just dropped MobileLLM-Pro on Hugging Face a 1B foundational language model in the MobileLLM series, designed to deliver high-quality, efficient on-device inference across a wide range of general language modeling tasks two variants of the model: A pre-trained base model https://x.com/_akhaliq/status/1978916251456925757
How Starcloud Is Bringing Data Centers to Outer Space | NVIDIA Blog https://blogs.nvidia.com/blog/starcloud/
Just to recap: We found out today that an LLM that fits on a high-end consumer GPU, when trained on specific biological data, can discover a novel method to make cancer tumors more responsive to immunotherapy. Confirmed novel discovery (not present in existing literature).”” / X https://x.com/deredleritt3r/status/1978561622932164905
Very excited to share @theworldlabs ‘s latest research work RTFM!! It’s a real-time, persistent, and 3D consistent generative World Model running on *a single* H100 GPU! Blog and live demo are available below! 🤩”” / X https://x.com/drfeifei/status/1978840835341914164
Introducing RTFM (Real-Time Frame Model): a highly efficient World Model that generates video frames in real time as you interact with it, powered by a single H100 GPU. RTFM renders persistent and 3D consistent worlds, both real and imaginary. Try our demo of RTFM today! https://x.com/theworldlabs/status/1978839171058815380
We are sharing a research preview of our latest model from @theworldlabs! RTFM is an autoregressive diffusion transformer trained on large-scale video data. It generates video frames in real-time without building an explicit 3D model of the world. Try the demo today!”” / X https://x.com/jcjohnss/status/1978842517605843391
Andrej Karpathy released nanochat, ~8K lines of minimal code that do pretrain + midtrain + SFT + RL + inference + ChatGPT-like webUI. It trains a 560M LLM in ~4 hrs on 8×H100. I trained and hosted it on Hyperbolic GPUs ($48). First prompt reminded me how funny tiny LLMs are. https://x.com/Yuchenj_UW/status/1978144157970661495
Qwen3-VL 30B-A3B at 4-bit precision, running on Apple silicon at 80 tok/s with MLX! @awnihannun @Prince_Canuma @ostensiblyneil @lmstudio https://x.com/vincentaamato/status/1977776546736713741
7. Small models can punch above their weight. Their best 4B-parameter model, trained with this recipe (real data + diverse RL + GRPO-TCR), beats 14B–32B models on tough benchmarks like AIME25 and GPQA-Diamond. Smart data and tuning trump raw size. Really good paper for AI devs”” / X https://x.com/omarsar0/status/1978112412743258361
Tiny Recursion Model (TRM) results on ARC-AGI – ARC-AGI-1: 40%, $1.76/task – ARC-AGI-2: 6.2%, $2.10/task Thank you to @jm_alexia for contributing TRM, a well written, open source, and thorough research to the community based on the HRM from @makingAGI https://x.com/arcprize/status/1978872651180577060
🚀 PaddleOCR-VL is here! Introducing PaddleOCR-VL (0.9B) — the ultra-compact Vision-Language model that reaches SOTA accuracy across text, tables, formulas, charts & handwriting. Breaking the limits of document parsing!🌍 Powered by: • NaViT dynamic vision encoder • ERNIE https://x.com/PaddlePaddle/status/1978809999263781290
Thanks to @NVIDIAAIDev, @xingyaow_ and I got to try an early NVIDIA DGX Spark system, which is this little GPU box that sits on your desktop. It’s sleek looking, fast, and quiet, but can easily run strong LMs like @Alibaba_Qwen 3 Coder, we wrote a little bit about using it:”” / X https://x.com/gneubig/status/1978067258506187238
since the day it was announced, i’ve been dying to get my hands on DGX Spark; a small but powerful machine i can put on my desk to run latest open models of almost any size. thanks to @nvidia, the dream came true a few weeks ago. look at this cutie sitting on my desk at NYU https://x.com/kchonyc/status/1978156587320803734
Thanks Jensen for the hand delivery of DGX Spark. Best delivery service ever. Amazing to see so much compute (1 petaflop!) in such a tiny form factor. https://x.com/gdb/status/1978273142695977391
Powering inference for the fastest growing AI companies like OpenEvidence, Writer, and Clay means being the first to use bleeding-edge model performance tooling in production. That’s why we were early adopters of NVIDIA Dynamo, giving us 50% lower latency and 60%+ higher https://x.com/basetenco/status/1978883986924634551
NVIDIA GTC Washington, DC, is just around the corner, happening October 27–29! I’ll be there in person, excited to join the Physical AI and Robotics sessions and, of course, Jensen’s keynote. Can’t wait to connect with many of you! https://x.com/TheHumanoidHub/status/1977777355197444308
These are the Qwen3-VL models I have most been looking forward to – in 4B and 8B sizes, I expect these will work well even on just CPUs”” / X https://x.com/simonw/status/1978151711987372227
ARM CEO Rene Haas says the embedded chip market for physical AI will be bigger than data centers. Today, robots largely use repurposed automotive chips, but these are not designed specifically for robot. There will be tens/hundreds of chips per bot – the market will be gigantic https://x.com/TheHumanoidHub/status/1978189401831174167
nanochat d32, i.e. the depth 32 version that I specced for $1000, up from $100 has finished training after ~33 hours, and looks good. All the metrics go up quite a bit across pretraining, SFT and RL. CORE score of 0.31 is now well above GPT-2 at ~0.26. GSM8K went ~8% -> ~20%, https://x.com/karpathy/status/1978615547945521655
One More (Small) Thing: Introducing mxbai-colbert-edge-v0 17M and 32M. They are are the result of an easily reproducible way to train ColBERT models from scratch. They’re strong, too: the 17M variant would rank first on the LongEmbed leaderboard for models under 1B parameters. https://x.com/mixedbreadai/status/1978853869557055492
Cognition | Introducing SWE-grep and SWE-grep-mini: RL for Multi-Turn, Fast Context Retrieval https://cognition.ai/blog/swe-grep
day 3 on nanochat and it’s getting integrated! – we now have weights for the small and large models with 20 and 32 layers. Karpathy shared his own checkpoints on the hub. – I’ve got demos for both weights on the hub as space. – we have a PR on transformers to integrate NanoChat, https://x.com/ben_burtenshaw/status/1978832914952401081
