An independent benchmark of on-device LLMs on iPhone A17 Pro found Apple's system model achieves ~149 tok/s with only 12MB peak app memory, while 4B-class open models like Qwen3 4B and Llama 3.2 3B tr…
Weill Cornell Medicine researchers published EmulatRx in Nature Communications on July 7, 2026, a five-agent LLM framework for clinical-trial design using real-world patient data from EHR datasets suc…
For the first time, a satellite has identified and described features in its image scans without needing to check back with ground control. The YAM-9 satellite uses a NASA-built program called NAVI-Or…
Researchers introduced a Bangla event detection benchmark with 9,979 annotated sentences across clean text, ASR transcripts, and corrupted text. They found encoder models perform better on clean text …
Parseflow, a new macOS app, extracts structured data from PDFs using a local AI model, ensuring 100% privacy with no cloud processing. Priced at €17.99 as a one-time payment, it offers unlimited usage…
A developer fully fine-tuned a 270M-parameter Gemma 3 model on a laptop using the Banking77 dataset, achieving ~96% accuracy on intent classification. The project used full fine-tuning with loss maski…
Loft Orbital's YAM-9 satellite became the first to run a vision-language model in orbit, enabling real-time classification of Earth imagery from plain-English queries. The satellite uses NASA JPL's NA…
SqueezLabs, a two-person company, has developed CrankGPT, a hand-powered chatbot that runs on a Raspberry Pi 5 with a hand-crank generator, designed for use in a post-apocalyptic scenario without grid…
On April 16, 2026, NAVI-Orbital became the first system to demonstrate a zero-shot vision-language model performing autonomous multi-modal inference entirely onboard a Low Earth Orbit spacecraft. The …
For the first time, an Earth observation satellite has used a vision-language model in orbit to autonomously identify areas of interest from natural language queries, without human analysts. The miles…
SambaNova Systems launched a European sovereign AI inference platform that offers managed, dedicated, and on-premises deployment options with full data residency in German infrastructure. The platform…
Engineers can now run a full local agentic programming stack using Ollama, Gemma 4, and Claude Code, bypassing API costs, rate limits, and data privacy concerns. Google DeepMind's Gemma 4 26B MoE mode…
Google DeepMind released Quantization-Aware Training (QAT) checkpoints for the Gemma 4 model family, targeting local deployment on edge devices and consumer GPUs. The new Q4_0 QAT format reduces memor…
Google Cloud demonstrated a multi-cluster inference setup deploying an LLM across two regional GKE clusters using TPU v6e accelerators and GKE managed DRANET for networking. The experiment, documented…
Researchers have developed Soro, a family of Tajik-specialized conversational AI models built from Gemma 3 checkpoints and trained on a 1.9-billion-token Tajik corpus. The models outperform same-size …
Financial Times testing, conducted in partnership with AI safety group Alice, found that safety controls in Meta's Llama 3.3 and Google's Gemma 3 open-weight AI models can be removed in under 10 minut…
A 4-billion-parameter AI model released in early 2025 is now outperforming models seven times its size on standard reasoning benchmarks, with Google's Gemma 3 4B scoring 89.2% on GSM8K math reasoning …