AI Open Source Radar — August 25, 2026

Selection criteria: AI repositories with the steepest daily star growth (stars today) over the last 24 hours. Repositories with large cumulative counts but flat or decelerating recent momentum — "rank-holders" — were excluded. Data: Trendshift daily, weekly and per-repository pages, cross-verified.

freestylefly/awesome-gpt-image-2 — +2.2k ⭐ today (14.8k total)

https://github.com/freestylefly/awesome-gpt-image-2

This is a "Prompt as Code" industrial prompt engine and template library built specifically for GPT-Image-2. The organizing idea is to compress prose-style prompts into structured protocols so that agents and scripts — not just humans — can reuse them, and the README states three explicit design principles:

Atomic schema — decompose subject, lighting, materials, layout and visual detail into composable parts
Workflow friendly — designed to be consumed by agents, scripts and automation systems
Structured control — improve controllability over layout, copy and information hierarchy

Roughly 520 cases have been reverse-engineered and sorted into 13 categories (Posters & Typography 82, Photography & Realism 75, UI & Interfaces 73, Illustration & Art 56, Charts & Infographics 52, Products & E-commerce 40, and more), alongside 20-plus industrial templates and a companion gallery site with large previews, full-prompt copy and in-browser test generation. Most recently the whole collection was extracted into an Agent Skill called gpt-image-2-style-library, installable into Claude Code or Codex with a single npx skills add command. The repository sat at #16 on the daily list on Aug 6, then reached #1 on GitHub Trending overall for the first time on Aug 23, kicking off a second wave — and today’s gain accounts for 92% of the entire week’s growth (+2.4k), confirming the spike is happening right now.

Practical angle: turns marketing and e-commerce imagery from one-off prompting into a template-plus-skill pipeline, enabling on-brand batch generation at volume.

#PromptEngineering #GPTImage2 #PromptTemplates #AgentSkill #DesignAutomation #Typography

FlashML-org/FreeToken — +852 ⭐ today (4.5k total) · second consecutive appearance

https://github.com/FlashML-org/FreeToken

Built by researchers at UC Berkeley and MIT, FreeToken is an edge-native Mixture-of-Experts inference engine that treats heterogeneous personal hardware — GPUs, CPUs, host memory, interconnects — as a single elastic serving platform. Its enumerated components are:

Bandwidth-adaptive CPU–GPU co-execution
Full-layer double-buffered prefill streaming
Global LRU expert caching
The FTW fast weight format
Semantic anchor checkpoints for recurrent state and KV caches
Dynamic runtime VRAM re-allocation between expert caches and KV memory with no engine restart or weight reload

The semantic anchors are the differentiator: they let agentic context edits — tool calls, inserted thinking blocks — avoid redundant recomputation, meaning the engine was shaped around agent workloads from the start. It serves both the OpenAI-compatible API (/v1/chat/completions, /v1/responses) and the Anthropic-compatible API (/v1/messages), so it drops straight into existing agent clients, and claims 20-plus MoE models spanning 35B on an 8GB laptop GPU, 284B on a gaming desktop, and 753B GLM-5.2 on a single workstation GPU. The arXiv paper (2608.16157) landed around Aug 17, then MarkTechPost on Aug 23 and Dataconomy on Aug 24 both ran the "753B on one GPU" hook — and with today’s gain exceeding the previous session’s +757, it is still accelerating.

Practical angle: lets clients with data-egress restrictions prototype frontier-class open-weight models on a single in-house workstation, sharply lowering the entry cost of an on-premise evaluation.

#MoEServing #EdgeInference #LocalLLM #VRAMOptimization #ExpertCaching #KVCache

zhaoxuya520/reverse-skill — +556 ⭐ today (27.3k total)

https://github.com/zhaoxuya520/reverse-skill

A security "skill router" pack for AI coding agents — explicitly not a tool installer, but a methodology-dispatch layer built on the principle "route first, execute second." When an agent encounters an APK, a binary, obfuscated frontend JavaScript, a packet capture, a CTF challenge or a pentest target, the pack first branches it to the correct methodology and sub-skill before any tool is touched. Mechanically, the agent executes Section 0 of an AI-only bootstrap file (README_AI.md) that detects the install path and OS (Windows, Kali, generic Linux, macOS), refreshes a machine-local tool index, then follows the global routing rules in RULES.md and the scenario-to-skill dispatch matrix in skills/routing.md through a canonical 14-step behavior chain. The 20-plus skill modules include:

apk-reverse, ida-reverse, js-reverse
binary-diff — cross-version symbol migration
patch-diff-exploit — N-day patch diffing into a working PoC
pwn-chain — stack, heap and kernel
firmware-pentest — following OWASP FSTM
edr-bypass-re and llm-security (OWASP LLM / ASI Top 10)

A separate CTF-Sandbox-Orchestrator adds 40-plus sub-skills. The "self-evolving" part lives in field-journal/, which forces the agent to write back experience on five triggers — task completed, new toolchain pitfall, bootstrap defect fixed, routing gap found, informative failure — and to read that index before starting anything new. After hitting #1 on GitHub Trending on Jul 31 it climbed steadily, and today’s gain represents 86% of the entire week’s growth (+648), a sharp re-acceleration. The project states use is limited to authorized security research, learning and CTF.

Practical angle: standardizes the "which methodology applies here" judgment that junior engineers most often miss, narrowing the quality variance across a security team’s penetration-test reports.

#ReverseEngineering #PentestAutomation #SkillRouting #BinaryDiffing #CTF #FirmwareAnalysis

wafer-ai/gpu-perf-engineering-resources — +293 ⭐ today (1.9k total)

https://github.com/wafer-ai/gpu-perf-engineering-resources

A single-file curriculum covering GPU kernel programming and AI infrastructure performance engineering from fundamentals through production deployment, produced by YC S25 startup Wafer under an MIT license. Its most distinctive feature is the reading order: every subsection is split into Tier 1, 2 and 3, and the recommended path is breadth-first — read all Tier 1 material across every topic before moving to Tier 2, rather than going deep on one subject. Eleven top-level sections, each with three subsections, cover fundamentals (GPU programming, architecture, low-level PTX/SASS), matrix multiplication (tutorials, advanced implementations, cuBLAS internals), tensor cores and mixed precision (including Blackwell), attention and memory-bound kernels (FlashAttention 1/2/3, PagedAttention, KV cache), compilers and DSLs (Triton, CUTLASS & CuTe), profiling and optimization, AMD and alternative hardware (ROCm, CDNA, TPU/Pallas), production inference systems (vLLM, SGLang, TensorRT-LLM, continuous batching, speculative decoding), LLM-generated kernels (KernelBench, KernelLLM, TritonBench, AI CUDA Engineer, AlphaEvolve), distributed and multi-GPU work (NCCL, parallelism strategies, kernel fusion), and industry analysis. Contribution rules are explicit: primary sources and practitioner blogs with real implementation insight only — no surface-level tutorials and no unverified AI-generated content. It carries no Trendshift badges and no prior trending history at all, making it the cleanest new entrant on the list; 15% of its cumulative stars arrived today, and 91% of the week’s gain (+322) landed in this single day.

Practical angle: usable as-is for onboarding a team tasked with cutting inference costs, giving the organization a shared vocabulary for diagnosing "why is our serving slow."

#CUDA #KernelOptimization #FlashAttention #TritonDSL #InferenceServing #TensorCores

Reading Today’s Board

Today’s four repositories attack the same problem from four different layers. FreeToken works beneath the hardware, the Wafer curriculum works on the people who operate that hardware, and reverse-skill and awesome-gpt-image-2 sit above the agent, supplying methodology and prompt assets respectively. The theme of the day, in other words, is optimizing everything around the model rather than the model itself.

Excluded: Large and Rank-Holding Repositories

deepseek-ai/deepseek-harness (+2.8k today, 189.6k total) — today’s gain doubled from the previous observation (+1.4k on Aug 23), but with 189.6k cumulative stars plus #1 Repository Of The Day and #1 Repository Of The Week badges, and having already been featured on Aug 18, standard procedure excludes it. The trigger was the rc.1/rc.2 releases on Aug 21 (the DeepSeek-V4-Flash-Vision-Exp multimodal model plus a Bubblewrap sandbox-escape security fix).
DietrichGebert/ponytail (+581 today, 109.7k total) — four separate #1 badges, continuously trending since Jun 12.
stablyai/orca (+639 today, 51.9k total) — #1 on GitHub Trending on Aug 11, trending since Jul 10, with no verifiable trigger event in the last three days.
tinyhumansai/openhuman (+298 today, 37k total) — #1 on GitHub Trending on May 11, with a documented nine-day run at the top.
VoltAgent/awesome-agent-skills (+339 today, ~31.8k total) — consistently high-ranked since December 2025.
Alishahryar1/free-claude-code (+516 today, 45.4k total) — remains excluded as a rank-holder.
Leonxlnx/unlazy — +184 in the previous session, +47 today; excluded under the decay rule.