Skip to content

Blog

Latest posts

RSS
  1. Published · on research.meta.ai

    A million-token window, weights included

    Meta released weights for Muse Spark Open: 816B parameters, a 1M-token context, and a stated path to run it on two Mac Studios.

  2. Published · on developer.nvidia.com

    One open model reads the receipt and the table

    NVIDIA's Nemotron 3.5 Super VL handles text, documents and images in one open model; Ramp uses it to turn receipts into JSON.

  3. Published · on alo-dlm.github.io

    ALoDLM spends more compute on the hard tokens

    Amazon AGI's ALoDLM gives each token as many recurrent passes as it needs, and decodes GSM8K about 2.7x faster than vLLM-served Qwen3-8B.

  4. Published · on deepmind.google

    Nano Banana 2.1 tops its own image editing scores

    Google's Nano Banana 2.1 leads its image editing benchmarks, with multi-character consistency at 1106 Elo, but factuality stays weak.

  5. Published · on microsoft.github.io

    CUAWright writes programs instead of clicking

    Microsoft's CUAWright gives browser and desktop agents a terminal, so a task becomes a script that can be rerun and verified.

  6. Published · on anthropic.com

    Haiku 5.5 cuts the price of a short request by 90%

    Anthropic's Haiku 5.5 costs 90% less than Haiku 4.5 for prompts up to 100,000 tokens, the band where most requests sit.

  7. Published · on github.com

    DeepGEMM Ascend runs DeepGEMM kernels on Huawei NPUs

    DeepSeek's DeepGEMM Ascend keeps the DeepGEMM API on Huawei Ascend NPUs, with dense GEMM reaching up to 99.8% of the hardware limit.

  8. Published · on allenai.org

    AstaBrief writes a cited report in 51.1 seconds

    Ai2 has open-sourced AstaBrief 8B, which turns a research question and literature excerpts into a cited report in about 51 seconds.

  9. Published · on cactuscompute.com

    Whistle transcribes speech from a 16.9 MB file

    Cactus Compute's Whistle is a 16.9 MB open speech recognition model that runs on CPU in the same C++ engine as its Needle model.

  10. Published · on blog.google

    Gemini 4 Argon raises output limit to 1M tokens

    Google's Gemini 4 Argon lifts the output token limit from 64K to 1M and goes first to trusted cyber defenders, not to enterprises.

  11. Published · on huggingface.co

    A 6B image model that outputs transparent backgrounds

    inclusionAI's Ming-Image-0.1-Design is a 6B text-to-image model for text-rich designs that returns RGBA images with transparent backgrounds.

  12. Published · on github.com

    OpenMuse puts an agent in a browser and a terminal

    OpenMuse is an alpha, self-hosted agent template that browses, runs commands and keeps a task record you can inspect.

  13. Published · on huggingface.co

    Ovis-VL-Embedding-9B puts text and documents in one search space

    A 9B vision-language embedding model maps text, images, documents and video into one space, so one index can serve mixed business content.

  14. Published · on aikido.dev

    Aikido shrinks a security model to run on your own servers

    Aikido's Altar shows a large model can be shrunk to fit customer infrastructure, the same trade-off office AI faces.

  15. Published · on anthropic.com

    Claude agents found a CRISPR-like enzyme system

    Anthropic reports that Claude agents found a previously uncharacterized enzyme system in bacteriophages; its function is still unknown.

  16. Published · on github.com

    Nimble returns typed decisions, not chat text

    A small open model that answers only from a fixed list of choices, with a probability for each, could fit invoice and approval checks.

  17. Published · on huggingface.co

    UK AISI shares benchmark results with setup details

    Benchmark scores now arrive with the setup details needed to judge whether two results were measured the same way.

  18. Published · on anthropic.com

    Anthropic announces Claude Opus 5.5 at lower cost

    Anthropic says Opus 5.5 matches its Fable 5.1 tier on most work while costing 40% less to run than Opus 5.