GitHub Trending Weekly 2026-04-22: Skills Ecosystem Explosion, Self-Evolving Agents Go Mainstream, Voice AI Dual Race
Data period: 2026-04-14 – 2026-04-22 (Rolling 7 days) Sources: GitHub Trending weekly + monthly, GitHub Search API, HN Algolia
TL;DR: Karpathy's CLAUDE.md derivative andrej-karpathy-skills topped the chart with +44,394 weekly stars by a landslide, igniting this week's biggest trend — a full-blown Skills framework ecosystem explosion, with 10+ skill-type repos appearing across both trending and new repo charts; NousResearch Hermes Agent crossed the 100K total star milestone, making "self-evolving agents" this week's second keyword; VoxCPM2 and Voicebox advanced on dual tracks, quietly laying the infrastructure foundation for open-source voice AI.
Fastest Growing — Weekly Star Gains Top 15
Source:
github.com/trending?since=weekly🔁 = Also appearing in monthly trends (sustained momentum signal)
| # | Project | +Stars/wk | Total Stars | Lang | Created |
|---|---|---|---|---|---|
| #1 🔁 | forrestchang/andrej-karpathy-skills | +44,394 | 71,863 | — | 2026-01-27 |
| #2 🔁 | NousResearch/hermes-agent | +30,630 | 108,034 | Python | 2025-07-22 |
| #3 | thedotmack/claude-mem | +12,472 | 65,121 | TypeScript | 2025-08-31 |
| #4 🔁 | microsoft/markitdown | +7,084 | 114,020 | Python | 2024-11-13 |
| #5 | multica-ai/multica | +7,009 | 18,471 | TypeScript | 2026-01-13 |
| #6 | jamiepine/voicebox | +5,936 | 22,127 | TypeScript | 2026-01-25 |
| #7 | Lordog/dive-into-llms | +5,703 | 33,329 | Jupyter | 2024-04-08 |
| #8 | EvoMap/evolver | +4,032 | 6,307 | JavaScript | 2026-02-01 |
| #9 | virattt/ai-hedge-fund | +3,950 | 56,811 | Python | 2024-11-29 |
| #10 | lsdefine/GenericAgent | +3,914 | 5,496 | Python | 2026-01-16 |
| #11 | BasedHardware/omi | +3,634 | 11,822 | Dart | 2024-03-22 |
| #12 🔁 | shiyu-coder/Kronos | +3,227 | 20,054 | Python | 2025-07-01 |
| #13 🔁 | OpenBMB/VoxCPM | +3,189 | 15,348 | Python | 2025-09-16 |
| #14 | openai/openai-agents-python | +3,078 | 24,360 | Python | 2025-03-11 |
| #15 | SimoneAvogadro/android-reverse-engineering-skill | +2,299 | 4,421 | Shell | 2026-02-02 |
Top New Repos — This Week's Newborns Top 15
Source: GitHub Search API (
created:2026-04-14..2026-04-22, sorted by total stars)
| # | Project | Total Stars | Lang | Created |
|---|---|---|---|---|
| #1 | kyegomez/OpenMythos | 6,690 | Python | 2026-04-18 |
| #2 | browser-use/browser-harness | 4,372 | Python | 2026-04-17 |
| #3 | Robbyant/lingbot-map | 3,875 | Python | 2026-04-15 |
| #4 | alchaincyf/huashu-design | 2,839 | HTML | 2026-04-19 |
| #5 | vercel-labs/wterm | 2,269 | TypeScript | 2026-04-14 |
| #6 | lewislulu/html-ppt-skill | 1,754 | HTML | 2026-04-15 |
| #7 | Nightmare-Eclipse/RedSun | 1,683 | C++ | 2026-04-15 |
| #8 | tw93/Kami | 1,413 | HTML | 2026-04-20 |
| #9 | cathrynlavery/diagram-design | 1,320 | HTML | 2026-04-16 |
| #10 | Manavarya09/design-extract | 1,272 | JavaScript | 2026-04-15 |
| #11 | codejunkie99/agentic-stack | 1,250 | Python | 2026-04-15 |
| #12 | EvoLinkAI/awesome-gpt-image-2-prompts | 1,131 | Python | 2026-04-18 |
| #13 | wbh604/UZI-Skill | 1,077 | Python | 2026-04-16 |
| #14 | BuilderPulse/BuilderPulse | 1,044 | — | 2026-04-14 |
| #15 | VoltAgent/awesome-claude-design | 1,006 | — | 2026-04-18 |
This Week's Spotlight — Fastest Growing Top 15
#1 🔁 — forrestchang/andrej-karpathy-skills | Karpathy's LLM Coding Principles Distilled into CLAUDE.md
A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls.
This week +44,394 ★ | Total ★71,863 | — | Created 2026-01-27
This is the week's most predictable yet most worth-dissecting viral event. Developer Forrest Chang distilled Andrej Karpathy's public observations on LLM coding pitfalls — over-engineering, ignoring existing patterns, pulling in unrequested dependencies — into a single CLAUDE.md file. The repo claimed the #2 spot for most new stars globally on April 13 alone, and topped this week's chart again with a +44K lead.
The real takeaway isn't the number itself, but the market signal it reveals: developer demand for "AI coding behavior standardization" has hit critical mass. A Markdown file accumulating 70K stars means engineers are willing to spend time configuring AI behavior boundaries rather than accepting defaults. That's a powerful positive feedback signal for the entire Claude Code Skills ecosystem.
This repo also appears in monthly trends (🔁), confirming this isn't a one-time burst but sustained discovery by newcomers.
#2 🔁 — NousResearch/hermes-agent | Self-Improving Agent Crosses 100K Total Stars
The agent that grows with you
This week +30,630 ★ | Total ★108,034 | Python | MIT | Created 2025-07-22
Hermes Agent crossed the 100K total star milestone this week, with +30,630 weekly gains still commanding second place. Nous Research's open-source flagship project's core proposition is closed-loop learning: agents generate reusable skills from each task, continuously refine them in subsequent use, and build persistent cross-session memory models of the user.
According to TokenMix.ai benchmarks, self-created skills can reduce research task time by 40% — but this figure is domain-specific, and cross-domain skill transfer remains an unsolved problem (something the Nous research team themselves acknowledge). A 108K-star repo with 6,010 open issues is a signal worth monitoring for serious adopters.
The team maintains a companion sub-project hermes-agent-self-evolution, which uses DSPy + GEPA frameworks to optimize skills, prompts, and code — currently one of the most concrete technical roadmaps for agent self-improvement in the open-source community.
#3 — thedotmack/claude-mem | Cross-Session Memory Plugin for Claude Code, 46K Stars
A Claude Code plugin that automatically captures everything Claude does during your coding sessions, compresses it with AI (using Claude's agent-sdk), and injects relevant context back into future sessions.
This week +12,472 ★ | Total ★65,121 | TypeScript | Created 2025-08-31
The timing of claude-mem's surge is telling — it arrived right after the andrej-karpathy-skills wave of "Claude Code configuration fever," suggesting a cohort of developers who entered the Claude Code ecosystem quickly ran into the memory persistence problem, then discovered this tool.
The technical approach is straightforward: five lifecycle hooks (SessionStart, UserPromptSubmit, PostToolUse, Stop, SessionEnd) automatically capture all Claude operations, compress summaries via Claude Agent SDK, store them in local SQLite + ChromaDB vector index, and auto-inject relevant context at the next session start. Installation is a single command: npx claude-mem install.
As of this week, claude-mem has accumulated 46K stars across 223 releases with 92 contributors. For long-term project developers, this is currently the most complete out-of-the-box memory persistence solution in the Claude Code ecosystem.
#4 🔁 — microsoft/markitdown | The 110K-Star Veteran Still Burning
Python tool for converting files and office documents to Markdown.
This week +7,084 ★ | Total ★114,020 | Python | MIT | Created 2024-11-13
markitdown is this week's only repo above 100K total stars that still maintains 7K+ weekly gains. As a preprocessing standard tool for the AI toolchain, it converts PDF, Word, Excel, PowerPoint, HTML, and other formats to Markdown, making them directly consumable by any LLM pipeline. Support for autogen, langchain, and other major frameworks keeps driving sustained interest.
Continued monthly trend presence (🔁) indicates markitdown has entered the "essential tool list for new developers entering AI development" — growth driven by compounding word-of-mouth, not events.
#5 — multica-ai/multica | Managing Coding Agents Like Full-Time Employees
The open-source managed agents platform. Turn coding agents into real teammates — assign tasks, track progress, compound skills.
This week +7,009 ★ | Total ★18,471 | TypeScript | Created 2026-01-13
multica's proposition: instead of directing agents with ad-hoc prompts, integrate them into your issue tracker — assign tasks, track progress, accumulate skills. Last week its HN title was "Your next 10 hires won't be human" (3 points, 2 comments) — no heated debate, but 7K weekly stars suggest developers are actually trying it rather than just watching.
Combined with this week's agentic-stack (new repo) and its portable .agent/ memory architecture, the narrative "agent as teammate, not disposable tool" is being validated simultaneously from multiple directions.
#6 — jamiepine/voicebox | Open-Source Voice Synthesis Studio, Powered by Qwen3-TTS
The open-source voice synthesis studio
This week +5,936 ★ | Total ★22,127 | TypeScript | MIT | Created 2026-01-25
Voicebox is one half of this week's "voice AI dual race." Positioned as an open-source Eleven Labs, it's built on Qwen3-TTS, Whisper, and MLX, offering voice cloning, real-time transcription, and voice design, running on CUDA or Apple Silicon.
The notable technology choice: using Qwen3-TTS instead of OpenAI's TTS API is a deliberate sovereignty statement — keeping voice capabilities local without paying cloud API fees. Read alongside #13's VoxCPM2: this week's open-source voice AI competition axis is "model architecture innovation (VoxCPM's tokenizer-free diffusion)" vs. "engineering integration completeness (Voicebox's studio UX)."
#7 — Lordog/dive-into-llms | Hands-On LLM Tutorial, 33K Stars and Still Going
《动手学大模型 Dive into LLMs》hands-on programming tutorial series
This week +5,703 ★ | Total ★33,329 | Jupyter Notebook | Created 2024-04-08
This Chinese-language LLM hands-on tutorial collection charts again with +5,703 weekly stars. Unlike most trending repos this week, dive-into-llms was last committed in October 2025, meaning growth is driven by long-term word-of-mouth rather than new feature attraction.
For engineers or learners entering the LLM space, the Jupyter Notebook format provides a complete practical path from basic fine-tuning to RLHF — still one of the most widely recommended resources in the Chinese-speaking learning community.
#8 — EvoMap/evolver | GEP-Driven AI Agent Genome Evolution Engine
The GEP-Powered Self-Evolution Engine for AI Agents. Genome Evolution Protocol.
This week +4,032 ★ | Total ★6,307 | JavaScript | GPL-3.0 | Created 2026-02-01
evolver introduces the "Genome Evolution Protocol" (GEP) concept — treating an agent's skills and strategies as mutable genomes, letting evolutionary pressure select the most effective combinations. On HN, two adjacent discussions (DuoRAG self-evolution, ShinkaEvolve) scored low (1-3 points), but 4K GitHub stars suggest developers have experimental interest even as the HN community remains cautious.
evolver, GenericAgent (#10), and hermes-agent (#2) together point to this week's implicit theme: "letting agents decide what to learn and how to evolve" has moved from research papers into installable open-source tools.
#9 — virattt/ai-hedge-fund | AI Hedge Fund Team, 56K-Star Financial Agent Framework
An AI Hedge Fund Team
This week +3,950 ★ | Total ★56,811 | Python | Created 2024-11-29
ai-hedge-fund uses multiple AI agents playing different roles — analysts, risk managers, decision makers — simulating real hedge fund operations. This week's +3,950 gains aren't event-driven, reflecting sustained search volume for financial AI frameworks in the quant community.
Read alongside #12's Kronos (financial market foundation model): the financial AI open-source ecosystem is advancing on two parallel tracks — upper-layer multi-agent workflows (ai-hedge-fund) and lower-layer prediction models (Kronos) are developing their own communities.
#10 — lsdefine/GenericAgent | Self-Evolving Agent That Grows a Full Skill Tree from 3.3K Lines
Self-evolving agent: grows skill tree from 3.3K-line seed, achieving full system control with 6x less token consumption
This week +3,914 ★ | Total ★5,496 | Python | MIT | Created 2026-01-16
GenericAgent makes a specific technical claim: starting from 3,300 lines of seed code, the agent builds its own skill tree, achieving full system control while consuming 6x fewer tokens than baselines. Topics include skill-tree, self-evolving, computer-control, memory-system — the most explicit technical statement in this week's self-evolution theme.
The 6x token efficiency claim hasn't been independently verified, so it's worth testing before deployment. However, with 3,900 weekly star gains and MIT license, it's a worthwhile starting point for developers who want to experiment with "minimum viable agent self-growth."
#11 — BasedHardware/omi | Wearable AI That Sees Your Screen, Hears Your Conversations
AI that sees your screen, listens to your conversations and tells you what to do
This week +3,634 ★ | Total ★11,822 | Dart | MIT | Created 2024-03-22
omi is this week's most actively discussed repo on HN (19 points, 13 comments) — higher than all other repos combined. It pairs wearable hardware (necklace/smart glasses) with a phone app for continuous conversation monitoring, screen observation, and real-time suggestions.
The core HN debate is the tension between privacy and utility: "Wait, it's always recording?" coexists with "I tested it, the call summary feature is genuinely useful." This controversy itself signals that omi hits a real user need rather than being just a tech novelty. The Flutter/Dart frontend with Python backend enables simultaneous iOS and Android deployment.
#12 🔁 — shiyu-coder/Kronos | First Open-Source Financial K-Line Foundation Model, AAAI 2026 Selection
Kronos: A Foundation Model for the Language of Financial Markets
This week +3,227 ★ | Total ★20,054 | Python | MIT | Created 2025-07-01
Kronos is the first open-source financial K-line (OHLCV) foundation model, pretrained on over 12 billion K-line records across 45 global exchanges. The paper has been accepted at AAAI 2026.
The technical highlight is a two-stage architecture: first using a specialized tokenizer to quantize continuous multi-dimensional K-lines into hierarchical discrete tokens, then pretraining a large autoregressive Transformer on these tokens. According to the paper, price sequence prediction RankIC improves 87% over the best non-pretrained baseline, and volatility prediction MAE drops 9%. Continued monthly trend presence (🔁) indicates the quant community is in sustained evaluation mode.
#13 🔁 — OpenBMB/VoxCPM | Tokenizer-Free TTS, the Voice AI Architecture Competition
VoxCPM2: Tokenizer-Free TTS for Multilingual Speech Generation, Creative Voice Design, and True-to-Life Cloning
This week +3,189 ★ | Total ★15,348 | Python | Apache-2.0 | Created 2025-09-16
VoxCPM2 from Tsinghua's OpenBMB team is an open-source voice generation model with the core proposition of a tokenizer-free architecture: using an end-to-end diffusion autoregressive architecture to directly generate continuous speech representations, skipping discrete tokenization, theoretically producing more natural and expressive synthesized speech.
Specs: 2B parameters, based on MiniCPM-4 backbone, trained on 2M+ hours of multilingual speech data, supporting 30 languages, outputting 48kHz audio. The "Voice Design" feature lets you generate entirely new voices using just natural language descriptions (gender, age, emotion, speaking rate) without reference audio.
Read alongside Voicebox (#6): two voice AI repos charting the same week signals that open-source voice AI infrastructure is accelerating — Voicebox takes the integration engineering route, VoxCPM takes the model architecture innovation route, and the two aren't mutually exclusive.
#14 — openai/openai-agents-python | OpenAI's Official Multi-Agent Framework, Steady Growth
A lightweight, powerful framework for multi-agent workflows
This week +3,078 ★ | Total ★24,360 | Python | MIT | Created 2025-03-11
OpenAI's official multi-agent framework maintained steady 3K weekly gains with no specific event driver. Compared to Hermes Agent's (#2) 30K gains, this gap reflects two fundamentally different growth patterns: "hot community repo" vs. "official authoritative tool."
openai-agents-python's competitive advantage is native official integration: mechanisms like Handoff, Guardrails, and Structured Outputs are all maintained by OpenAI, making it suitable for production environments that depend on stable API guarantees.
#15 — SimoneAvogadro/android-reverse-engineering-skill | Claude Code Android Reverse Engineering Skill
Claude Code skill to support Android app's reverse engineering
This week +2,299 ★ | Total ★4,421 | Shell | Apache-2.0 | Created 2026-02-02
This Shell Skill enables Claude Code to perform Android APK static analysis, decompilation, and manifest parsing. 4.4K total stars for a highly specialized security research tool is notable, indicating Claude Code Skills adoption is extending from general development into niche verticals like security research.
This week's Skills theme exploded from the trending chart (#1, #15) to new repos (huashu-design, html-ppt-skill, UZI-Skill, diagram-design, agentic-stack, awesome-claude-design), confirming "extending AI coding agent capabilities via skills" has become the developer community's core action direction this week.
This Week's Spotlight — Top New Repos Top 10
New #1 — kyegomez/OpenMythos | Open-Source Reconstruction of Claude Mythos, 770M Params Matching 1.3B
A theoretical reconstruction of the Claude Mythos architecture, built from first principles using the available research literature.
Total ★6,690 | Python | MIT | Created 2026-04-18 HN: 6 points, 2 comments
OpenMythos by Kye Gomez (creator of the Swarms framework) garnered 6,690 stars within 4 days of release. Its core hypothesis: Claude Mythos belongs to the "Recurrent-Depth Transformer" (RDT) architecture — rather than stacking hundreds of unique layers, a set of layers loops multiple times, with the entire inference process occurring in continuous latent space via a single forward pass.
MarkTechPost-verified efficiency claim: a 770M-parameter looped model, on the same training data, matches the downstream performance of a 1.3B fixed-depth Transformer — halving the parameter count. This is a "theoretical reconstruction," not "official documentation" — Anthropic has never publicly disclosed Mythos's actual architecture — but its rigor is sufficient to warrant serious attention from researchers.
New #2 — browser-use/browser-harness | Removing the Framework Layer, Letting LLMs Drive Chrome CDP Directly
Self-healing browser harness that enables LLMs to complete any task.
Total ★4,372 | Python | MIT | Created 2026-04-17 HN: 3 points, 1 comment
browser-harness comes from the browser-use team (authors of the original browser-use framework), but this time the direction is completely reversed: removing the framework. Instead of wrapping CDP with Playwright APIs, it gives the LLM a raw Chrome DevTools Protocol connection and a helpers.py — when steps fail, the agent reads errors, self-edits helpers, and retries. That's the "self-healing" mechanism.
browser-use founder Gregor Zunic wrote on X: "We got tired of browser frameworks restricting the LLM. So we removed the framework." This philosophical shift caught the community's attention — is it a better solution, or does it just shift complexity to the LLM? Early results suggest it works well with models that have strong reasoning capabilities, while weaker models may produce more unpredictable behavior.
New #3 — Robbyant/lingbot-map | Feed-Forward 3D Scene Reconstruction from Streaming Data
A feed-forward 3D foundation model for reconstructing scenes from streaming data
Total ★3,875 | Python | Apache-2.0 | Created 2026-04-15
LingBot-Map is the most technically distinctive new repo this week — it's not an agent, not a skill, not an LLM tool, but a geometry-aware 3D foundation model. It can reconstruct 3D scene geometry in real-time from streaming image sequences using a feed-forward architecture (non-iterative), meaning inference speed outpaces traditional NeRF-class methods.
This technology has direct applications in autonomous driving, AR/VR, and robot navigation — running on a parallel track to this week's AI agent/skills wave. 3,875 total stars indicate the computer vision community is paying attention.
New #4 — alchaincyf/huashu-design | HTML-Native Design Skill with 20 Design Philosophies
Huashu Design · HTML-native design skill for Claude Code
Total ★2,839 | HTML | Created 2026-04-19
huashu-design represents the design vertical in this week's Skills explosion. It gives Claude Code HTML-native high-fidelity prototyping capabilities — 20 design philosophies, 5-dimension review framework, 31 layout types, 20+ animation effects, with MP4 export. The author is alchaincyf, well-known in the AI coding community (creator of nuwa-skill and other popular skills).
The "agent-agnostic" design means it works not just with Claude Code but in any environment supporting skills frameworks. This direction — designers distilling their design expertise into reusable agent skills — is one of the most noteworthy expansion trends in this week's Skills ecosystem.
New #5 — vercel-labs/wterm | Zig + WASM-Powered Browser-Native Terminal Emulator
A terminal emulator for the web
Total ★2,269 | TypeScript | Apache-2.0 | Created 2026-04-14
wterm is the strongest candidate for "best-engineered new repo this week." Vercel Labs wrote the core VT100/VT220/xterm escape sequence parser in Zig, compiled to a ~12KB WASM binary, and handed the rendering layer back to DOM — native text selection, browser find, and accessibility come free.
The philosophy behind the tech stack: don't reinvent terminal parsing in JS; use Zig's memory control and zero-cost abstraction for the core; let the browser's native rendering do what browsers do best. A React wrapper (@wterm/react) minimizes integration cost. For developers who need a real terminal embedded in a web app, this is the cleanest architecture option available.
New #6–#10 — Skills Ecosystem Second Tier
The back half of this week's new repo chart features five skills and tools, each targeting a different vertical:
lewislulu/html-ppt-skill (1,754 ★): HTML PPT generation Skill with 24 themes, 31 layouts, 20+ animations — a "no PowerPoint needed" presentation workflow.
cathrynlavery/diagram-design (1,320 ★): 13 types of editorial diagrams, pure HTML + SVG, deliberately avoiding Mermaid ("No shadows, no Mermaid-slop"). Ideal for generating high-quality visualizations in agent workflows.
Manavarya09/design-extract (1,272 ★): Extract any website's complete design system with one command, outputting DTCG tokens with support for iOS SwiftUI, Android Compose, Flutter, Figma variables, shadcn/ui, and more. MIT licensed with an MCP server for direct Claude Code workflow integration.
codejunkie99/agentic-stack (1,250 ★): Portable .agent/ folder (memory + skills + protocols) supporting Claude Code, Cursor, Windsurf, Hermes Agent, and other coding harnesses. The "keep your knowledge when switching tools" design direction is this week's most direct response to the agent portability problem.
wbh604/UZI-Skill (1,077 ★): Stock analysis Skill with 51 investing master perspectives, 22 data dimensions, 180 quantitative rules, and 17 institutional analysis methods, supporting A-shares, HK stocks, and US stocks. Named after the Chinese trading slang "UZI" (ultra-fast attack trading strategy) — this week's most regionally flavored Skill.
Monthly Trend Cross-Reference
Repos overlapping with monthly trends this week (🔁 marked):
- andrej-karpathy-skills (#1): Sustained monthly momentum, confirming the Karpathy CLAUDE.md effect spans multiple weeks as newcomers continuously discover and star it upon entering the Claude Code ecosystem
- NousResearch/hermes-agent (#2): Sustained monthly momentum, Hermes with 100K total stars is now the benchmark open-source agent project of 2026
- microsoft/markitdown (#4): Monthly evergreen, AI preprocessing toolchain essential continues compounding via word-of-mouth
- shiyu-coder/Kronos (#12): Sustained monthly, quant community in ongoing evaluation of the AAAI-accepted quantitative foundation model
- OpenBMB/VoxCPM (#13): Sustained monthly, voice AI researchers and developers tracking VoxCPM2's progress
Three of the five monthly overlaps are AI agent/tools (andrej-karpathy-skills, hermes-agent, markitdown), one is financial AI (Kronos), and one is voice AI (VoxCPM) — perfectly mapping to this week's three main narratives.
This Week's Trend Insights
1. Skills Ecosystem Goes from Niche to Mainstream: Claude Code Configuration Culture Is Official
This week's chart features over ten repos that explicitly identify as skills or Claude Code plugins, spanning security research (android-reverse-engineering-skill), design (huashu-design, diagram-design), presentations (html-ppt-skill), investing (UZI-Skill), and multi-platform portability (agentic-stack). This is no longer early-adopter territory — when a CLAUDE.md file can top the chart with 44K weekly gains, "writing behavior specs for AI coding agents" has become a routine action for mainstream developers. The result of configuration culture taking hold is that AI tools no longer have just one out-of-the-box usage mode.
2. Self-Evolving Agents Move from Research to Installable Tools
hermes-agent (#2), GenericAgent (#10), and evolver (#8) all charted the same week, meaning "letting agents decide what to learn and how to evolve" has moved from research papers into pip install-able open-source tools. These three projects take different technical approaches (Nous's closed-loop learning, GenericAgent's skill tree seed, evolver's GEP genome evolution), but all point to the same future: agents are no longer static tools but partners that grow with use.
3. Infrastructure Week for Open-Source Voice AI
Voicebox (#6) and VoxCPM2 (#13) both charted the same week, alongside ai-hedge-fund and Kronos forming a financial AI dual track — the open-source ecosystem is clearly accelerating in two "non-agent" domains simultaneously this week. The voice AI competition axis is now clear: VoxCPM targets model architecture innovation via tokenizer-free design (theoretically higher performance ceiling), while Voicebox targets engineering integration completeness (immediately usable). For most developers, Voicebox has a lower entry barrier; for scenarios requiring custom voice design or multilingual support, VoxCPM2 offers greater flexibility.