Category
Startup Underdog released the open model Woof (32 MB, 31 million parameters) for speech generation and earlier the model Bark (167 MB) for speech transcription. Both run locally on the device without an internet connection.
Researchers have proposed the ACSeL method, which searches for an LLM configuration (model size, batch, quantization) for operation on edge devices with limited memory and latency. Tests were conducted on an NVIDIA RTX A6000 and Jetson Thor.
The Phonon TTS speech synthesis model runs directly on the device without internet or cloud. The smaller Phonon 31m version (65 MB, English) has the Apache 2.0 license, usable commercially as well, while the Phonon 40m version (75 MB, 5 languages) with voice cloning is under CC BY-NC.
Hugging Face fine-tuned the 20B open-weights model Palette for creating presentations in the IBM Design Language. The model splits the work into a planning role and a coding role and, according to the company, took first place in two out of three domains of the PresentBench benchmark.
Kenyan startup Fastagger has integrated the offline AI assistant Auni for small businesses into the M-Pesa app; over 6000 businesses have signed up since its launch in 2024. The assistant analyzes sales and works even on cheap phones without internet.
In the Cheeky Pint podcast, Sam Altman said it is "terrible" that DRAM sits in personal computers utilized at only 5%, while the world faces a memory shortage. He advocates for its maximum utilization in the cloud.
An employee of NVIDIA created the MonsterHouse Halloween installation – a talking garage door playing the role of a grumpy pirate ship. It uses the Nemotron Voice Agent pipeline with the Nemotron ASR, Bolt E4B, Magpie TTS, and Audio2Face-3D models, running offline on a single workstation with RTX PRO cards.
The University of Turin published a textbook-style book on LLMs on arXiv for scientists and professionals without programming knowledge, including a practical exercise on local deployment using Ollama and Open WebUI.
New packages for Apple platforms allow developers to choose a speech transcription model, distinguish between speakers, and obtain typed responses with probabilities. Open models can be downloaded from Hugging Face Hub and run directly on the device.
GitHub has made experimental automatic selection between supported local and cloud models available. The feature requires users to opt in, and its availability depends on the application and operating system. Routing decisions continue to be made in the cloud.
A preview version of WebAI is available in the Opera One browser for running AI models directly on the device. Users must enable the feature. According to Opera, it supports Gemma models as well as the local foundation model from Apple on Mac computers; performance depends on the hardware.
The study authors deployed the LLENS platform within the closed network of Hokkaido University Hospital. It runs on eight NVIDIA H200 GPUs with the Kimi K2.6 model. When working with hospital documents, they measured accuracy of 97.2 % with an interactive approach compared with 81.3 % with fixed retrieval.
Ollama now supports local text-to-speech using Phonon models on macOS with Apple Silicon. It offers streaming playback and an API compatible with the OpenAI API. Licensing varies by model variant.
The PyTorch 2.15 release candidate expands MPS support on Mac computers to include FP8. The team at Hugging Face reports up to 6.5 times faster MoE model decoding and a reduction in peak memory usage during video decoding from 270 to approximately 8 GiB.
AnyLanguageModel 1.0 for Swift has been released. The same session code supports the system model, local models and hosted services. According to the authors, tool calling and structured generation work across all providers. The library also runs on Linux.
Bidirectional KV cache transfer in vLLM allows earlier computations to be reused in subsequent conversation turns. In its own measurements, AWS reports time-to-first-token speedups of up to 1.6-fold for Qwen3-32B and up to 3.3-fold for zai-org/GLM-5-FP8.
On RTX Spark laptops, Hermes, Perplexity and OpenClaw agents can be installed through the Windows onboarding experience. Windows Store handles updates. However, the article’s author encountered an error while setting up the OpenClaw agent that required reinstallation.
Apple presented a procedure for running Muse Spark Open locally using MLX and RDMA. According to its own measurements, it achieved a speed of over 36 tokens per second on four Mac Studio computers. The model has 816 billion parameters, of which 42 billion are active.
Starting with version 1.0.94-0, GitHub Copilot CLI discovers supported local models from a running Ollama service using the /model command. Once confirmed, they can be used without restarting. Selecting a local model does not enable offline mode or disable telemetry.
The study presents CourseChat, an assistant that allows six separate courses to share two local AI servers. The authors compared models and retained the production model with 8 billion parameters. The system offers 435 prepared questions; the study does not demonstrate improved learning outcomes.
AI Radar monitors Czech and international sources every day, looking for changes that truly deserve attention.
The result is a quick overview of what has actually changed in the AI world, rather than another stream of articles.
Use the CS/EN switch to read the same Radar in Czech or English. English content is published after its translation has been checked, so new and older items may appear later.
Each row represents one event — not one article. At a glance, you can see its significance, credibility and main point.
Illustrative example, not a current news item.
Agent mode is available to all paying users
Until now, the mode was available only on the highest plan; it is now available on all paid tiers without a waitlist.
The detail page contains a fuller summary, its significance and original sources. Practical impact appears in the detail and the For individuals and For businesses views. An AI Flash item reaches the main selection only after it has been expanded and meets the publication rules.
We first summarize each event in the same way for everyone. Based on those same facts, we then explain what the change means for your own use and what it could mean for how a company operates.
Today’s briefing is the same for everyone. Pages For individuals and For businesses can be found in the main navigation — they select only events relevant to the given use case.
Reset history · Outage history · Subscribe in a feed reader (RSS) · Methodology and label explanations
