Open source

Stanford Open-Sources Memorilla: Latent Memory for Frozen LLMs

Stanford researchers open-sourced Memorilla, a memory system that compresses documents into 16 tokens for frozen LLMs. It achieves 84.33% accuracy on the FactKG benchmark, using 108x fewer document tokens than full context.

Summarized by AI from one outlet

Source

Latest

  1. Open source

    datasette-llm supports OpenAI-compatible endpoints

    Simon Willison released datasette-llm 0.1a11, adding support for OpenAI-compatible endpoints. The update allows Datasette users to configure LLMs hosted locally via LM Studio or Ollama, utilizing a unified web protocol interface to enhance development toolchain flexibility.

    Simon Willison

  2. Open source

    Cmpnd Releases dspy-wsp: Weight-Space Programming Comes to DSPy

    Cmpnd released dspy-wsp, a prototype DSPy optimizer enabling frontier LLMs to read and edit small model weights. The method requires no gradients and costs about $5 to $10 per run, offering a new approach to model tuning.

    MarkTechPost

  3. Open source

    DeepSeek V3 rewritten in ~200 lines of pure Jax

    A developer released an open-source project on Hacker News that refactors the DeepSeek V3 model using approximately 200 lines of pure Jax code. The project demonstrates a concise implementation of the architecture within the Jax framework, offering a new experimental basis for researchers.

    Hacker News

  4. Open source

    Cloudflare Open-Sources Kitesurf AI Browser

    Cloudflare has open-sourced Kitesurf, a browser designed for AI agents that runs on Cloudflare Workers without requiring Chromium processes. It supports the Chrome DevTools Protocol and claims significantly lower CPU and memory consumption compared to Chromium, optimizing for agent tasks such as HTML extraction and screenshots.

    IT之家

  5. Open source

    Runway Open-Sources Robotics Model Praxis-1

    Runway released its open-source robotics model Praxis-1 under the Apache 2.0 license. The model leverages video training knowledge for robot control and leads the World Action Model category on Nvidia's RoboLab-120 benchmark, though it cannot yet plan complex action sequences independently.

    THE DECODER

All news → · All rights belong to the original creators