MarkTechPost
·
1 day ago
PrismML released Bonsai 27B, a quantized version of Qwen3.6-27B using 1-bit and ternary weight compression. The ternary variant achieves 5.9GB model size while retaining 94.6% of FP16 baseline performance, and the 1-bit variant reaches 3.9GB with 89.5% retention. These models enable running 27B-class quality inference on laptops and phones with practical memory constraints and improved throughput on resource-limited devices.
Simon Willison
·
1 day ago
A developer created a custom animated pet for Codex Desktop by using GPT-5.6 Sol and gpt-image-2 to generate sprite assets for a pelican riding a bicycle. The AI system required multiple rounds of image generation and refinement, starting with a detailed reference prompt that specified dimensions of 192x208 pixels and a magenta chroma-key background for animation. The resulting assets—including sprite sheets and animation loops—are now published in the developer's GitHub repository, with the underlying image generation and pet creation skills available as open-source Apache 2.0 licensed code.
The New Stack
·
1 day ago
Enterprises are increasingly concerned about protecting their data from AI labs while facing high inference costs from closed-source models like OpenAI and Anthropic, leading some to propose using lower-cost open-source models trained on private data instead. Major AI labs including OpenAI, Anthropic, and xAI have stated that API usage is not used to train their models and some offer zero data retention plans, but enterprises remain skeptical of these assurances. As a result, companies like Microsoft are considering alternatives such as open-weight Chinese models or homegrown models to reduce dependency on AI labs that are simultaneously competing in software categories where enterprises operate.
NVIDIA
·
2 days ago
NVIDIA Nemotron Labs promotes open AI models that enterprises can customize and control for domain-specific tasks, contrasting with closed proprietary models. Companies like Harvey achieved legal task accuracy matching frontier models at 10x lower cost, while Arcee AI reached inference costs of approximately 90 cents per million tokens, roughly 20x cheaper than comparable closed models. This shift enables organizations to build specialized AI applications tailored to their specific workflows and data rather than adapting their needs to existing general-purpose models.
Simon Willison
·
2 days ago
Datasette released version 1.0a37, continuing its development toward a stable 1.0 release. The project has accumulated 1,526 monthly briefing subscribers as of July 14, 2026. Simon Willison is seeking sponsorship at $10 per month to fund continued development and curated LLM news digests.
The New Stack
·
2 days ago
Open-source and open-weight AI models are claimed to be approximately four months behind proprietary frontier models while costing roughly ten times less per token, with Featherless demonstrating annual costs of $90,000 for their optimized GLM 5.2 model versus $1.5 million for GPT-5.5 or Claude Opus at 100 billion monthly tokens. The cost comparison shows that enterprises paying for proprietary models are primarily paying for enterprise wrapper features like integration and observability rather than raw model intelligence. As open-source models improve and approach frontier capability levels, enterprises may find themselves locked into expensive proprietary contracts while open alternatives become increasingly viable alternatives.
The Neuron
·
2 days ago
Osaurus is a macOS application that runs open-source AI models locally on Apple Silicon Macs without sending data to external servers. The software is free, MIT-licensed, and allows users to integrate ChatGPT, Claude, or Gemini for specific tasks while maintaining a shared memory across all models. Users can build autonomous agents with voice control, file monitoring, and browser extensions that operate offline and independently.
MarkTechPost
·
2 days ago
Hayden Bleasel released Blume, an open-source documentation framework that converts Markdown folders into production documentation sites without configuration. The tool generates a hidden Astro project, ships version 1.0.3 on npm, and requires Node.js 22.12 or newer. Blume includes built-in AI features such as llms.txt support, per-page Markdown export, and a Model Context Protocol server for Claude and other tools to read documentation directly.
