AI Open Source Radar — August 23, 2026

Method: ranked primarily by stars gained today (Trendshift daily, as of 2026-08-23 13:42 UTC), cross-checked against weekly momentum. Cumulative stars are shown for reference only.

Today is a complete changing of the guard. The DeepSeek Harness ecosystem that dominated mid-August has collapsed to +1.4k a day (down 90% from its August 18 peak), and all five repos featured on August 21 — openbot, ip-as-logo-skill, OpenViking, fx and maka — have dropped out of the daily top 25. Four new entries take their place. Note also what is not here: mattpocock/skills (+1.0k today), openai/codex (+1.3k) and free-claude-code (+291) all posted big daily numbers, but each has been trending since April with cumulative star counts in the tens or hundreds of thousands — rank-holders, not risers.

FlashML-org/FreeToken — +757 ⭐ today (2.9k total, 196 forks)

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

FreeToken is an edge-native Mixture-of-Experts serving engine built to run frontier-scale open-weight models at interactive speed on the gaming PC you already own, without datacenter GPUs. Its core idea is to stop treating GPU, CPU, host memory and interconnect as separate resources and instead present them as one elastic inference pool. Four components make that work:

Bandwidth-adaptive CPU–GPU co-execution
Full-layer double-buffered prefill streaming
Global LRU expert caching
Graph-compatible execution paths

Using official checkpoints without extreme quantization, the project reports a 35B-class MoE running on an 8GB laptop GPU, 284B on a gaming desktop, and the 753B GLM-5.2 on a single workstation card, with support for DeepSeek-V4-Flash, Qwen3.6-35B-A3B and GLM-5.2. The differentiator is that it exposes Anthropic- and OpenAI-compatible APIs, so existing agent tooling like Claude Code or Codex can switch to a local model by changing an endpoint rather than any code. Why now: the FreeToken-desktop build shipped with one-click Windows and Linux installers and a GUI — no GGUF conversion, no building from source — and “my 8GB 4060 is running a 35B model” posts cascaded across X, landing the same week the arXiv paper (2608.16157) went up. Python, Apache 2.0, 3 contributors.

Practical angle: for clients blocked from cloud LLMs by token budgets or data-residency rules, you can keep their existing agent toolchain intact and repoint the endpoint at an in-house workstation to deliver frontier-class models on-premises.

Tags: #EdgeInference #MoEServing #CPUGPUOffload #ExpertCaching #ConsumerGPU #OnPremLLM

HiThink-Tech/Financial-API — +227 ⭐ today (1.3k total, 79 forks)

https://github.com/HiThink-Tech/Financial-API

This is the official A-share (mainland China equities) data service published by HiThink Tonghuashun, one of China’s largest financial data providers, exposing real-time quotes, historical candles, financial statements, indices, sectors, limit-up moves, unusual activity, hot lists and Dragon-Tiger block-trade disclosures behind a single API key. Structurally, the important design choice is that it ships four access paths at once:

REST API
MCP server
CLI
Python SDK (and an agent-skill packaging)

Register it as an MCP server in a coding agent like Claude Code or Codex and the model can translate a plain-language request — “pull the financial metrics for every stock that hit limit-up in the last five sessions” — directly into data queries. The primary use cases are individual quant research and agent-driven investment research, and the differentiator is provenance: retail quants have historically stitched together unofficial third-party APIs for A-share data, and here the exchange-grade provider opened an official channel. That is exactly why it is spiking — “the official data finally speaks MCP” spread through Chinese-language quant and AI communities over two days. TypeScript, MIT.

Practical angle: in a finance-domain agent proof of concept, registering one MCP server replaces building a market-data pipeline, compressing demo timelines from weeks to days.

Tags: #AShareMarketData #QuantResearch #MCPServer #RealtimeQuotes #FinancialStatements #FintechAgent

Leonxlnx/unlazy — +184 ⭐ today (1.3k total, 72 forks)

https://github.com/Leonxlnx/unlazy

unlazy is an agent skill that structurally prevents coding agents from cutting corners and declaring “done” — the cluster of failures researchers call model laziness, underthinking and premature completion. Its core technique is the Depth Tree: split a task into N layers, then give every leaf node the full time budget of the entire original task, so effort multiplies rather than divides as depth increases. On top of that, each leaf carries its own acceptance criteria, and the agent must record the output of a verification command as evidence — nothing counts as complete until the check passes. Inside Claude Code, a Stop hook physically blocks the agent from ending the session early. The design philosophy — replace “please try harder” prompting with verifiable hard constraints on what completion means — is grounded in 2025-2026 research on laziness and underthinking, which is what drove simultaneous pickup across Chinese- and English-language X yesterday. JavaScript, MIT, 2 contributors.

Practical angle: drop it onto long refactors or large migrations — the jobs where an agent most easily narrows scope mid-run — to enforce output review gates at the skill layer rather than in your own code.

Tags: #DepthTree #TaskDecomposition #StopHook #VerifiableAcceptance #PrematureCompletion #LongHorizonAgent

cclank/lanshu-create-ai-presenter-video — +172 ⭐ today (479 total, 31 forks)

https://github.com/cclank/lanshu-create-ai-presenter-video

This is a Codex skill that turns one script plus one authorized photo of a presenter into a finished, quality-checked vertical talking-head video. The pipeline runs end to end in a single pass:

Narration scripting
Speech synthesis
Digital-human generation
Lip-sync alignment
Subtitles and motion graphics
Editing and cover image
A quality-inspection report

Output is fixed at 9:16, 1080×1920, 30fps, 45–75 seconds. The biggest design differentiator is that it is provider-neutral: it is not bound to any particular video or speech API but picks up whatever tools exist in the environment, so swapping models or vendors requires no code changes. It installs into Cursor, Codex, OpenCode and other coding CLIs with a single npx command and runs entirely locally. Why now: a textbook post-launch explosion — it cleared 80 stars within hours of release, then the author’s own walkthrough video circulated, and 36% of its 479 cumulative stars arrived today alone. Python, MIT, 1 contributor.

Practical angle: use it as a weekly production line for repetitive short-form video — product update announcements, internal training — where updating the script is the only manual step and no camera or editor is needed.

Tags: #DigitalHuman #LipSync #TextToSpeech #VerticalVideo #CodexSkill #ProviderNeutral

The Thread Running Through Today

Today’s board converges on one theme: routing around a resource constraint by design. FreeToken works around a VRAM ceiling, unlazy works around a model’s shortfall of effort, lanshu works around the absence of a production crew, and Financial-API works around restricted data access.