Release
sno-ai/sno-station v1.5.0 — Reranking memory search results
sno-ai/sno-station v1.5.0 reranks memory search results, moving the memory that matches the answer to the top with a small ranking model.
CURATED OPEN SOURCE INTELLIGENCE
Latest news
Release
jimoto-no-llm/rustdsh v0.7.0 — Debian/Red Hat package distribution
jimoto-no-llm/rustdsh v0.7.0 is a release that broadens installation paths through Debian and Red Hat packages plus additional package manager manifests.
New
diegosouzapw/omniroute v3.8.51 — multi-provider AI gateway
A look at the multi-provider AI routing features of diegosouzapw/omniroute and the limits of the material on its latest release.
New
doronp/gitmemory v0.1.0 — Git-based agent memory retrieval
doronp/gitmemory v0.1.0 is an LLM-free memory tool that preserves agent conversations in Git and retrieves them by byte offset.
New
professorpalmer/bonsai-ada-surgery bundle-20261017b — a 12GB serving recipe
A repository documenting a recipe for serving Bonsai 2 27B on a 12GB RTX 4070, along with a hierarchical KV cache and tool-grammar optimization.
Release
abacusai/simless v0.5.2 — Fixes missing hosted tests
abacusai/simless v0.5.2 fixes a problem where, because of how package test results were ordered, "simless test all" passed even though the app's hosted test targets were missing.
New
jundot/omlx v0.7.1rc1 — Local Inference Server for Apple Silicon
jundot/omlx is a local inference server that runs LLMs on Apple Silicon and is managed through a macOS menu bar app.
New
rockmatrix-ai/edgerock — Constant-Memory Multimodal Inference Research
A research repository targeting the growing key-value cache problem in long inference, releasing inference and evaluation code plus the EdgeRock-4B model.
New
tongyi-mai/opengcua v0.1.0 — a code-execution computer use agent stack
tongyi-mai/opengcua is a general-purpose computer use agent stack that combines GUI manipulation, accessibility tree reading, app API calls, and shell commands as code; its latest release is v0.1.0.
Release
code-yeongyu/oh-my-openagent v5.1.35 — Typecheck and sync fixes
code-yeongyu/oh-my-openagent v5.1.35 fixes type package references, memory sync merging, and categories configuration error guidance.
New
qwenlm/blenderworld — a Blender 3D benchmark for visual agents
qwenlm/blenderworld is a benchmark that evaluates visual agents turning real-world references into editable Blender scenes.
New
flashml-org/freevideo — local video inference engine
flashml-org/freevideo is an inference engine that runs MiniMax H3 locally to generate video with ComfyUI.
New
ruvnet/ruflo v3.56.10 — Agent harness for Claude Code and Codex
Summarizes ruvnet/ruflo's agent orchestration features, per-client installation paths, and integration scope.
Release
davidondrej/cloudroom-core core-v0.1.62 — Idle agent control fix
davidondrej/cloudroom-core 0.1.62 fixes the retry of control commands for idle cloud agents and the related behavior.
New
nabzx/mnemosyne v0.0.8 — Agent Memory Version Control
nabzx/mnemosyne is a Rust project that brings version control and provenance tracking to agent memory; its latest release is v0.0.8.
New
receptron/laya v0.2.0 — Jev-compatible decision model without Python
receptron/laya is a Jev-compatible System-1 decision model that runs on Node.js and TypeScript through ONNX Runtime, and its latest release is v0.2.0.
Release
dongnguyenvie/bashcut v0.0.31 — Batch/Export Structure and Subtitle Fixes
dongnguyenvie/bashcut v0.0.31 is a macOS video editor release that tidies up the batch processing and export structure and fixes behavior in subtitles, search, and more.
New
hkuds/nanobot v0.3.5 — a self-hosted personal AI agent framework
Introduces the features and verifiable scope of hkuds/nanobot, a Python-based self-hosted personal AI agent framework.
New
rzc156/vla-diag — A diagnostic tool for VLA evaluation
rzc156/vla-diag is a Python diagnostic tool that checks what a VLA benchmark's success rate actually measures.
New
kvmem/kvmem v0.18.0 — host-memory KV virtualization for inference
kvmem/kvmem v0.18.0 is a C++ project that keeps the KV cache in host memory and brings it back to the GPU when needed, for long-context inference.
No matching stories.