Open Source
-
Research
Anthropic says Claude wrote machine-checked proofs that 123 open-source libraries cannot crash on their fuzz tests, working directly on compiled code
In a project with the verification startup Theorem, Anthropic had Claude produce Rocq proofs covering 239 fuzzing entry points in 123 OSS-Fuzz libraries at the RISC-V binary level, patching hundreds of bugs it hit along …
-
Open Source
Anthropic hands Agent Skills to the Agentic AI Foundation, which also takes in projects from OpenAI, Hugging Face and Alibaba
Anthropic has transferred the Agent Skills standard, including its name, GitHub organization and agentskills.io domain, to the Linux Foundation's Agentic AI Foundation, one of four projects the foundation accepted on Oct…
-
Open Source
Ai2 releases OlmoEarth v1.3 with open embeddings for the whole planet and a no-login explorer, claiming it beats Google's AlphaEarth on most tests
The Allen Institute for AI's new satellite embedding model turns a year of Sentinel and Landsat imagery into a 128-number vector for every 10-meter pixel, ships precomputed yearly maps of Earth under an open license, and…
-
Open Source
NVIDIA open-sources Flare, a Rust-based local inference server for RTX Blackwell workstations and DGX Spark aimed at coding agents and long conversations
NVIDIA published Flare on GitHub on October 20, 2026 under the Apache 2.0 licence. The research prototype prepares models with PyTorch tooling and serves them from a Rust runtime without Python, targeting SM120 and SM121…
-
New Models
IBM releases Granite Speech 5.0 2B and upgrades two companion models, open speech recognition under Apache 2.0
IBM's Granite team published granite-speech-5.0-2b on October 21, 2026, a multilingual transcription and translation model, along with revised Turbo CTC and 4.1 2B Plus models. IBM says the new 2B model ranks second amon…
-
Open Source
DeepSeek open-sources DeepInfer, the production inference engine behind DeepSeek-V4.1-Flash, with support for NVIDIA and Huawei chips
DeepSeek published DeepInfer, the in-house software it uses to serve DeepSeek-V4.1-Flash, under the Apache 2.0 license on October 20, 2026. The release runs on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, but leaves out…
-
Research
Anthropic says Claude flagged 493 attack-surface issues in 25 open-source projects and helped OpenSSH wall off pre-login code
Anthropic's Frontier Red Team reported on October 20, 2026 that Claude mapped code reachable earlier than intended across 25 large open-source projects and flagged 493 issues. In OpenSSH, the work led maintainers to ship…
-
Infrastructure & Chips
llm-d reaches version 1.0, giving companies an open-source control plane for running LLM inference across GPU and TPU fleets
The llm-d project released version 1.0 on October 20, 2026. Backed by Red Hat, IBM, Google Cloud, CoreWeave and NVIDIA and hosted by the CNCF, the Kubernetes-based inference stack adds multi-tenant flow control, router f…
-
Open Source
Google open-sources TorchTPU, making its TPUs a native PyTorch device built with Meta, Hugging Face, vLLM and SGLang support
Google released TorchTPU as open source on October 20, 2026 at PyTorch Conference. The native PyTorch backend, developed with Meta's PyTorch team, lets existing models train and serve on Google's TPUs with few code chang…
-
New Models
LightOn releases SlideOn and mSlideOn, open sparse retrieval models built from its dense and late-interaction retrievers
French AI company LightOn on October 19, 2026 released SlideOn, a 174M-parameter English sparse retriever, and mSlideOn, a 339M-parameter multilingual model for long documents and code, with weights, sparse autoencoder c…