Open-Source AI

Today's Stories

1. 1

   Chatbot on Your Own PC: Privacy, Offline, Free

   Numerous large language models (LLMs, AI that understands and generates text) from big names like Meta and Google can now be downloaded for free and run locally on your own computer, with no subscription fees or usage limits. Running an LLM locally means offline access and greater privacy, because you are not sending anything to the cloud for others to analyze. You also avoid paying AI companies monthly and gain a more personalized, private AI system, though you will handle updates yourself and models tend to be less advanced than paid ones.

   For a smooth experience, having plenty of RAM helps—8 GB is the bare minimum, but 16 GB is better and 32 GB or more is needed for the biggest models. On Windows, a dedicated Nvidia GPU with over 8 GB of VRAM makes a significant difference.

2. 2

   Meta open-sources Muse Glimmer, keeps Muse Spark 1.2 closed

   Meta Platforms released Muse Glimmer, a 30-billion-parameter AI model with open weights, on Aug. 10. As of Aug. 25, it had not released the weights for its more powerful flagship model, Muse Spark 1.2. Unlike OpenAI and Anthropic, Meta doesn't rely on subscriptions or APIs to monetize AI models. Instead, it uses AI to improve ads and recommendations on Facebook and Instagram, which already showed results: in Q2, AI improvements boosted Facebook ad clicks by 8.3% and ad conversions by 15.7%.

   Meta's capital spending for 2026 is now expected to be $130 billion to $145 billion, up from the initial $115 billion to $135 billion. The company also has $81.6 billion in noncancelable commitments due in 2027 and another $68 billion in data center leases starting in 2027-2028.

3. 3

   LAION releases 10M-hour open video dataset

   LAION released the Big Video Dataset (BVD), one of the largest open video datasets for AI research. It includes 10 million hours of video, 55 million clips with auto-generated descriptions, and 300 million still images. Models trained on BVD outperform comparable models trained on InternVid by up to 2.1 percentage points on common video-to-text benchmarks. The dataset is available for research only, and LAION's legal basis may rest on a 2024 Hamburg court ruling allowing copyrighted content for non-commercial research.

   The dataset and code are freely available, but users are asked to respect creators' rights. The data draws from 80 million downloaded videos out of 1.3 billion URLs found in CommonCrawl, mostly from YouTube and predominantly in English.

4. 4

   Nvidia's reported $13B Hugging Face deal highlights open-weight AI M&A surge

   Nvidia is reported to be acquiring Hugging Face for $13 billion, following its $6 billion deal with Poolside and Stripe's $7 billion acquisition of OpenRouter. These deals reflect a trend of tech giants hedging their bets on open-weight AI models, which are increasingly used for cost-efficient inference workloads, despite current low adoption rates of just 6% among companies.

   Fireworks CEO Lin Qiao, whose company processes 40 trillion tokens a day, predicts every company will eventually build its own specialized models, signaling a shift toward open-model adoption as frontier lab prices rise.

5. 5

   Hugging Face unveils $399 duck robot Microduck

   Hugging Face's French subsidiary Pollen Robotics announced the bipedal robot Microduck on August 27 and opened reservations at an introductory price of $399. Microduck is positioned as a platform for learning action-based AI, unlike the conversational Reachy Mini. Its small size (25cm tall) makes it safe and affordable for home or classroom use, with open-source software for retraining.

   Shipments are expected before Christmas 2026, initially in North America, Europe, and the UK, with four color options. The software stack, including SDK and learning scripts, is open-sourced under Apache 2.0.

6. 6

   Anthropic previews MHS standard for AI-operated machines

   Anthropic PBC previewed the Model Hardware Standard (MHS), a unified interface that lets AI agents control scientific instruments such as microscopes. It was developed with HHMI and is currently available only to a limited number of partners. MHS replaces device-specific APIs, which are often incompatible, with a single set of configuration commands. This makes it easier for researchers to automate lab workflows and allows AI agents to manage machines, as shown by HHMI's AI agent that operates microscopes and QuEra Computing's use of MHS with Claude to coordinate lasers.

   Anthropic plans to release MHS under an open-source license and says it could also control manufacturing equipment. Amazon Web Services, Hugging Face, and several industrial robot suppliers are already working to integrate MHS, while Anthropic will develop safety features with early adopters before broad availability.