Open-Source AI
Hugging Face, GitHub, Ollama, local models, agent frameworks
The lead
Latest first
Latest in Open-Source AI
One INT8 kernel swap changes every vLLM output
A new arXiv preprint swaps CUTLASS for Triton inside vLLM and finds no matching sequences, tracing the drift to scale rounding, not…
Agentao lets hosts veto every AI agent action before it runs
Agentao proposes a local-first runtime that splits an LLM's action proposals from host-authorized execution, adding permissions, replay, and audit trails.
arXiv paper scores AI harnesses without labels
New arXiv framework judges if an LLM-with-memory agent improves by measuring how fast a student converges to a stronger teacher, no labeled…
New Transformer swaps global attention for block memory
An arXiv paper shows BCMT matching dense Transformer quality on long context while raising training throughput and lowering memory use.
Batch pruning keeps reasoning models fast at scale
A training-free pruning method lets reasoning models keep accuracy and speed under batched inference, beating prior art by 39.7 points at 50%…
Microsoft patches 10 .NET flaws in August 2026 update
Microsoft's August 2026 .NET servicing update fixes 10 CVEs across .NET 8, 9, and 10, including remote code execution and privilege flaws.…
arXiv agent-memory model bars stale, retracted data
Researchers propose Governed Persistent Memory, an auditable model that blocks agents from releasing claims built on retracted, deleted, or stale records.
GitHub asks if your project is ready for AI-first contributors
GitHub's post asks whether open-source projects are ready for contributors who write and review code with AI. Here's what maintainers should weigh.
arXiv study: edge agents cut latency violations to 3%
Researchers propose MAS-DecStream, an edge scheduler where agents negotiate offloading with language models, cutting latency violations to 3% on the Alibaba trace.
LLMs collapse when asked to follow 7+ rules at once
An arXiv benchmark testing 15 models on up to 12 simultaneous rules finds reliable instruction following collapses past 5-6 constraints, falling below…