Top GitHub Projects: LLM Memory Compression Tool, AI Code Review Plugin, and WeCom CLI

TurboQuant Plus: Faster, Lower‑Memory Local LLM Inference

Running large language models locally is limited by VRAM, especially on Apple M‑series chips.

TurboQuant Plus implements the KV‑Cache compression described in the ICLR 2026 paper *TurboQuant*. The compression reduces KV‑Cache memory usage by 3.8‑6.4× while keeping quantisation accuracy close to Q8_0.

Turbo2 / Turbo3 / Turbo4 – three compression levels selectable according to available memory.

Sparse V‑gate attention decoding – 22.8% speedup for 32 K long‑context inference.

Layer‑adaptive mode 2 – different layers use distinct compression strategies.

Norm Correction – compensates numerical bias introduced by compression.

llama.cpp integration with Metal GPU kernel – native support for Apple M‑series GPUs.

Usage with llama.cpp:

# Load a model with TurboQuant compression
./main -m model.gguf --kv-cache-type turbo4

Python API example:

from turboquant import TurboQuantConfig

config = TurboQuantConfig(
    compression="turbo4",
    sparse_v_gate=True,
    norm_correction=True
)

Repository:

https://github.com/TheTom/turboquant_plus

Codex Plugin for Claude Code: AI‑Driven Code Review

Claude Code excels at code understanding, while OpenAI Codex excels at code generation. The plugin combines them so that Claude analyses source files, identifies issues, and invokes Codex to generate fixes or perform complex refactors.

Claude‑driven Codex execution – Claude detects problems and calls Codex for repair.

Complex task delegation – large refactors or cross‑file changes are handed to Codex.

Unified agent collaboration workflow – seamless handoff and data transfer between the two AIs.

Installation and configuration:

# Install the plugin
npm install -g @openai/codex-plugin-cc

# Enable in Claude Code
claude config set codex.enabled true
claude config set codex.api_key <your_openai_api_key>

Run a review of the current directory:

# Review code
claude review

Repository:

https://github.com/openai/codex-plugin-cc

WeCom CLI: Command‑Line Interface for Enterprise WeChat

Previously, interacting with the Enterprise WeChat Open Platform required the web admin console or custom API code. The CLI provides direct terminal access to the WeChat API.

CLI access to Enterprise WeChat API – invoke endpoints without writing code.

Message management – send messages, view history, manage groups.

Address‑book management – query departments, members, bulk import/export.

Application management – install apps, configure permissions, view statistics.

Rust implementation – cross‑platform (Windows, macOS, Linux) with low overhead.

Installation:

# macOS / Linux
curl -fsSL https://wecom-cli.github.io/install.sh | sh

# Windows
winget install WecomTeam.wecom-cli

Configure credentials:

wecom config set corp_id "YOUR_CORP_ID"
wecom config set agent_id "YOUR_APP_ID"
wecom config set secret "YOUR_APP_SECRET"

Example commands:

# Send a message to a user
wecom message send --user "Zhang San" --content "Working late tonight?"

# List departments
wecom department list

# Get member details
wecom user get --id "zhangsan"

# Broadcast a message to a department
wecom message broadcast --department "Tech" --content "Team building Friday afternoon"

Repository:

https://github.com/WecomTeam/wecom-cli