Token pooling-aware training cuts vector storage while keeping search quality
A new training method teaches multi-vector encoders to preserve retrieval signals after compressing document vectors, reducing storage by 75% with minimal quality loss.
Daily coverage of AI, developer tools and infrastructure. Each story explains what happened and why it matters, with a link to the original source.
A new training method teaches multi-vector encoders to preserve retrieval signals after compressing document vectors, reducing storage by 75% with minimal quality loss.
Meta AI released Muse Ember 1.0, a small model optimized for agent tasks and complex reasoning, priced competitively against GPT-6 Luna and Claude Haiku 5.5.
vLLM improves LLM inference throughput by replacing static batching with continuous batching and using PagedAttention to manage memory efficiently.
A community patch restores cuBLAS speeds for K-quant models on Kepler GPUs, fixing a significant slowdown caused by missing dequantization kernels.
MiniMax has released MiniMax-H3-Image, an open-source diffusion transformer that generates 16-megapixel images and edits references in a single checkpoint using six sampling steps.
Latent Space benchmarks Naive-N0.5-Flash against DeepSeek and GLM, finding high token throughput but variable reasoning costs.
Amazon Bedrock now hosts Kimi K3.1 Pro Highspeed, offering faster output speeds and cost-saving tiers for agentic workflows.
Anthropic releases Claude Fable 5.2, surpassing Gemini 4 Carbon on key benchmarks through extensive self-verification and subagent usage, despite higher per-task costs.
Google DeepMind researchers used the AlphaProof Code framework to mathematically prove the security of Signal's actual implementation, a task that previously took years.
A new iOS vulnerability allows apps to read other apps' private data, putting crypto keys at risk. Apple has not yet confirmed a fix in its latest updates.
Meta released Triton Low-Level Extensions (TLX), a new programming model that gives engineers explicit control over warp-group orchestration on modern GPUs without losing Triton's high-level productiv
NVIDIA Snapshot is an open-source tool that checkpoints initialized GPU memory on Kubernetes, restoring inference replicas in under 45 seconds.
A new open-source tool intercepts file reads and commands to hide API keys and personal data from AI agents like Claude Code and Codex, replacing them with placeholders.
NVIDIA introduces NVKE and NVKR, two open-source projects that help AI factories build and operate managed Kubernetes services on GPU infrastructure.
Git 2.56.1 and earlier maintenance releases fix CVE-2026-111760, an integer truncation bug that allows malicious repositories to execute code during recursive cloning.