Open-Source AI

Today's Stories

1. 1

   AWS releases Strands Harness, runs on any cloud

   AWS released "Strands Harness." Using Fable 5, it costs 77% less than Claude Code and also beats it on the Terminal Bench 2.1 score. Even with the same model, costs and performance appear to vary depending on the framework, giving developers an option when deploying agents to the cloud.

   The performance evaluation comes from AWS's own blog, with no third-party verification presented. It can be downloaded from GitHub or installed via pip or npm, and reproducibility in real-world operation will be the focus.

2. 2

   OpenFold3 production deployment speeds up antibody-antigen inference roughly 5x

   MindWalk announced the results of its production deployment of the OpenFold3 structure prediction model. Faster antibody-antigen work shows that the computing side could become a standard path for onboarding new enterprise partners.

   This result is specific to antibody-antigen inference, and it has not been confirmed that other prediction types such as proteins, RNA, DNA, and ligands would also see the same 5x speedup.

3. 3

   Anthropic, OpenAI safety claims go unverified, critics say

   A LessWrong essay argues that when Anthropic and OpenAI publish safety or alignment research, it is often entirely empirical, closed-source, and sparse on methodological details, leaving claims unverified. The essay says labs — or soon, their agents — can claim alignment progress that no one independently verifies, and it expects this kind of meta-science to remain systematically neglected.

   The proposed fix hinges on a dedicated effort to replicate alignment experiments, stress-test methodology, and open-source replications so outside researchers can validate and audit lab methods.

4. 4

   Base Browser ships Firefox hard fork with no AI slop

   The Base Browser project, hosted at codeberg.org/basebrowserproject, released a hard fork of Firefox that it labels as having no AI slop. A hard fork means the browser's code is maintained separately from Firefox, so users who want a browser without AI features may see this as an alternative, though the project's scale and support are unclear.

   The fork's viability hinges on whether the project can keep up with Firefox's security and feature updates over time; watch the project's activity on codeberg.org/basebrowserproject.

5. 5

   Alibaba's Qwen-Image-2.1: 7 billion parameters, beats most closed models on internal benchmark

   Alibaba's Qwen AI team released Qwen-Image-2.1, an open-weight model with a 7-billion-parameter visual component that runs on capable consumer GPUs like a 3090. It generates and edits transparent (RGBA) images and handles up to ten reference images at once. The team claims it beats most closed models on Qwen's own benchmark, though independent benchmarks are still pending. If verified, it suggests image generation and editing quality may be achievable on hardware creative professionals already own, rather than only through cloud-based closed services.

   The research license bars commercial use, so business adoption hinges on whether Qwen grants a separate license. Independent benchmark results will show whether the claim holds beyond Qwen's internal testing.

6. 6

   Tencent's Gander interrupts users in just 8 percent of cases, beats GPT-Realtime

   Tencent's Hunyuan Speech team and university researchers introduced Gander, which splits real-time conversation from background tasks between a "cerebellum" and a swappable "brain." On Full-Duplex-Bench v3, Gander interrupted users in 8 percent of cases versus 13.5 percent for GPT-Realtime. Gander's lower interruption rate suggests splitting conversation from reasoning can make voice assistants feel less disruptive than systems like GPT-Realtime, though its task accuracy fell slightly below that of the weakest competitor.

   The researchers call the work early, and how to scale Gander up remains an open question. Tencent plans to publish the weights and training data once it completes the open source release process.