AI Open Source Radar — August 2, 2026

Window: last 24 hours (UTC 2026-07-31 21:08 to 2026-08-01 21:08). Local (Korea) publication date: August 2, 2026. Primary signal is recent growth (findarepo 7-day measurement refreshed 2026-08-01, cross-checked with the OrangeBot daily 2026-07-31 snapshot). Cumulative stars are a reference only; large, perennially top-ranked repos are excluded.

drumih/turbo-fieldfare — +2,700 ⭐ (7 days) · ~3,000 ⭐ total

https://github.com/drumih/turbo-fieldfare

TurboFieldfare is a local inference runtime that runs Google’s Gemma 4 26B-A4B on Apple Silicon Macs using only about 2 GB of RAM. Instead of loading the full 14.3 GB model into memory, it keeps just the shared core (~1.35 GB) and the FP16 KV cache resident, then streams only the experts each token needs from SSD — letting even an 8 GB Mac run a 26B MoE model. The runtime, streaming installer, CLI and native Mac app are all written in Swift and Metal, and it is optimized specifically for this model rather than being a wrapper around MLX or llama.cpp. On first launch it downloads only the required byte ranges from Hugging Face, never materializing the full original checkpoint, and repackages them directly into a .gturbo format. Because Gemma 4 26B-A4B is an MoE model that activates only ~3.88 billion of its 26 billion parameters per token, this expert-streaming approach fits it especially well.

Practical use: Run a 26B-class model fully offline and local on a mainstream MacBook with no high-end GPU — ideal for privacy-sensitive on-device inference.
Why it’s surging: Nearly all of its stars arrived in the last seven days — a fresh breakout, as the promise of running a large MoE model locally on an affordable Mac spreads through the local-LLM community.

Tags: #AppleSilicon #MoE #ExpertStreaming #Gemma4 #LocalLLM #MetalGPU #OnDeviceInference

unicity-aos/aos-ce — +1,400 ⭐ (7 days) · ~8,600 ⭐ total

https://github.com/unicity-aos/aos-ce

AOS Community Edition is an open ‘agent operating system’ in which agents and agent-native software run. It composes models, tools, memory, skills and frontends as independently swappable ‘capsules,’ each with explicit authority boundaries that enforce least privilege. Sensitive actions are anchored on the Unicity blockchain via ‘Unicity Audit’ for a tamper-evident audit trail. Capsules are general user-space building blocks that can be assembled into harnesses, meta-harnesses, connectors or services, and it ships a construction tool called ‘Forge’ so a fresh agent can inspect the running system, find a real capability gap, and build and verify a least-privilege capsule itself. The aos CLI, HTTP API, distributions, first-party capsules, provider/model experience and Unicity Audit are all implemented in Rust.

Practical use: Serves as a governable agent runtime for enterprises that need fine-grained control over an agent’s file and tool access plus an audit log.
Why it’s surging: With rising interest in agent security and governance, its differentiator — an open agent OS combining explicit authority boundaries with on-chain auditing — is drawing fresh attention.

Tags: #AgentOS #CapsuleArchitecture #LeastPrivilege #RustLang #AgentGovernance #OnChainAudit

digimata/quill — ~+600 ⭐/day · 2-day streak · ~3,300 ⭐ total

https://github.com/digimata/quill

Quill is an ultra-minimalist, fully local macOS meeting recorder and transcriber. It records your mic and all system audio as two separate tracks, and when you stop it transcribes both on-device and writes a speaker-tagged transcript. Splitting the tracks serves two purposes: speech models are more accurate on clean single-source audio, and ‘me vs. them’ becomes free two-party diarization without any speaker-identification model. It runs as a single Swift binary with a menu-bar tray and no app bundle, using macOS 15+ Core Audio process taps to capture system audio — no virtual audio device or kernel extension required. The default transcription engine is Parakeet TDT 0.6B v2 (English) via FluidAudio’s Core ML port, processing about an hour of audio in roughly 20 seconds on Apple Silicon.

Practical use: A local-only alternative to cloud meeting tools like Otter and Granola — transcribe and separate speakers for sensitive meetings without sending anything off the machine.
Why it’s surging: Stars kept flowing in for a second straight day, as its cloud-zero, single-binary, privacy-first approach continues to draw interest.

Tags: #MeetingTranscription #SpeakerDiarization #CoreAudioTaps #ParakeetTDT #OnDeviceASR #SwiftBinary

AminBlg/SimpleEnglish — findarepo new tracking · ~1,100 ⭐ total

https://github.com/AminBlg/SimpleEnglish

SimpleEnglish is an agent skill that forces LLMs to write documentation in ASD-STE100 Simplified Technical English, the controlled language aerospace has used since 1983 so a tired mechanic cannot misread an instruction. For procedural text it applies rules like imperative mood, a maximum of 20 words per sentence and one instruction per sentence; for descriptive text, simple tenses, a maximum of 25 words per sentence, one topic per paragraph and at most six sentences per paragraph. It works in any harness that speaks the Agent Skills standard — Claude Code, Cursor, VS Code Copilot, OpenAI Codex, Gemini CLI, Goose, OpenCode and 25+ more — and ships as a single dependency-free folder under the MIT license. Across six Claude models and eight writing tasks, it produced 72.9% fewer STE violations per 100 words than the baseline, and output tokens dropped on every model.

Practical use: When writing manuals, API docs or safety instructions with an LLM — where misreading is costly — it quantifiably reduces ambiguity and ‘AI slop.’
Why it’s surging: As concern about ‘AI writing quality’ grows, its concrete result — porting an aerospace standard into a skill that cuts violations 72.9% — is spreading, driving a fresh climb.

Tags: #SimplifiedTechnicalEnglish #ASD-STE100 #AgentSkills #TechnicalWriting #ControlledLanguage #DocsAutomation