Cactus shipped Cactus Hybrid, a post-trained build of Google's Gemma 4 E2B that returns a calibrated confidence score with every answer, routing queries to a larger model when confidence falls below 0…
Google's Gemma 4 model uses a 68k-parameter probe layer that predicts decoding errors by reading hidden states, achieving a 0.79-0.88 AUROC on audio benchmarks despite being trained on zero audio data…
Apple's M5 generation introduces a dedicated Neural Accelerator on every GPU core, and BaseRT, a native Metal inference runtime, exploits these units to deliver up to 6.4× higher prompt-processing thr…
Samsung's Galaxy Z Fold 8 and Flip 8 are the first devices to launch with Gemini Intelligence and Gemini Nano 4, Google's latest on-device AI model based on Gemma 4 that supports over 140 languages wi…
Poolside AI's Laguna XS model, a small local LLM, performed well in refactoring Rust code via OpenCode, making relatively few errors and handling tool calls correctly except for escaped quotes. The 20…
Google announced that its Gemini Nano 4 model, now running on over 140 million devices, can be used through ML Kit's Prompt API to build on-device intelligent features in Android apps, such as summari…
Google's Gemma 4 E2B model serves efficiently on a single TPU v6e chip, achieving 213 tok/s for a single user and scaling to ~2,200 output tok/s across concurrent streams, while its QAT variants fail …
A developer has created tpu-management, a Claude Code skill and MCP server that automates the full lifecycle of serving Gemma 4 on Google Cloud TPUs, from zone selection and VM creation to benchmarkin…
Mira Murati's Thinking Machines Lab released Inkling, a 975B-parameter open-weight model, on July 15. Within hours, Artificial Analysis scored Inkling at 41 on its Intelligence Index, making it the to…
A single 24GB GPU like the RTX 3090 or RTX 4090 is the practical floor for serious local inference in 2026, and the best strategy is to run modern 20B–35B-class models at Q4_K_M quantization rather th…
A developer building an AI agent that controls Android phones using Gemma 4 discovered that the agent was misreading financial data 20% of the time due to a single bad assumption: treating OCR output …
A developer successfully ran Google's Agent Development Kit (ADK) with the locally hosted Gemma 4 model via Ollama, building a weather agent that uses tool calling to orchestrate Google Maps and Open-…
Thinking Machines has released Inkling, a 975B-parameter open weights model with 41B active parameters, debuting at 41 on the Artificial Analysis Intelligence Index and becoming the leading open weigh…
Google shipped a stealth update to its open AI model Gemma 4 that fixes tool calling bugs and truncated responses, while speeding up performance on Nvidia Hopper GPUs by 25 to 70 percent with Flash At…
A developer successfully ran Google's Gemma 4 E2B and E4B models on Apple's Neural Engine (ANE) using CoreML, achieving a memory footprint under 1 GB on an iPhone 17 Pro. The project overcame ANE's la…
A developer built a local MCP server for codebase memory using Ollama and ChromaDB, testing mistral:7b and ornith:9b models. The local mode runs entirely on-device, addressing privacy and billing conc…
A 26-billion-parameter Gemma 4 mixture-of-experts model can run at reading speed on a dual-socket 2013 Ivy Bridge Xeon server with no GPU, achieving ~5.2 tokens per second decode on a Q8_0 quant, acco…
A developer running Google's Gemma 4 26B mixture-of-experts model on a 13-year-old HP StoreVirtual server with dual Xeon E5-2690 v2 CPUs and no GPU achieved approximately 5 tokens per second inference…
Google announced Gemma 4 E2B for TPU, a model optimized to run natively on the Pixel 10's Tensor G5 TPU, at I/O Connect India. The multimodal model enables offline AI chat, image identification, and a…
Google unveiled Gemma 4 E2B for TPU and Functional Gemma at Google I/O India, demonstrating how Google Tensor's custom SoC and TPU enable 100% private on-device AI for the Pixel 10 family. The models …