At0mic News
Mistral's new platform lets enterprises train custom frontier models on their proprietary data. Unlike generic LLMs, Forge builds models that understand internal terminology, compliance policies, and workflows. Already partnered with ASML, European Space Agency, and others. Agent-first design with autonomous fine-tuning capabilities.
Major challenger to LM Studio. Web UI that trains 500+ models 2x faster with 70% less VRAM, runs locally on Mac/Windows/Linux, supports GGUF/vision/audio. Side-by-side model comparison, auto dataset creation from PDFs, code execution for testing. Apache licensed and llama.cpp compatible.
GPT-5.4 mini scores within 5% of full GPT-5.4 on coding benchmarks but 2x faster than GPT-5 mini. 400K context, multimodal. Nano optimized for simple tasks at 20 cents per 1M tokens (vs 75 cents for mini). Targeting live conversational workflows and task routing.
Core optimization work resuming after earlier setbacks.
Microsoft's "unhackable" 2013 console cracked via voltage glitching.
Researchers achieve 18-66x speedup in transformer learning with simple 5-line technique.
New version of the ubiquitous media processing toolkit.
Agent harness with planning tools, filesystem backend, and subagent spawning. 1,415 stars today.
Run Node.js apps inside WebAssembly sandboxes for better security.