JANGQ
JANGQ improves quantization of large language models by classifying tensors into sensitivity tiers and applying different bit widths: higher precision (5-8 bits) for attention layers and aggressive compression (2-4 bits) for MLP layers. This results in smaller models with superior performance compared to uniform MLX baselines, as demonstrated in curated benchmarks showing large MMLU gains with reduced memory. The open-source project includes runtime, profiles, and detailed results for models like MiniMax, Qwen, and Mistral.
Open SourceCLI
218stars
26forks
5features
2026since
Overview
5 featuresJANGQ is an AI & ML project. Standard uniform quantization degrading performance in smaller LLMs by treating all layers with equal precision regardless of sensitivity. JANGQ is an open-source project aimed at machine learning engineers and researchers. The project is open source (Open Source). JANGQ is available on the command line.
Behind JANGQ is JANG, and it first shipped in 2026. The project is developed in the open on GitHub with 218 stars and 149 commits in the last 90 days. Among its 5 catalogued features are Variable Bit Quantization, Layer Sensitivity Analysis, and MLX Optimization. The interface is available in 5 languages, including English, Spanish, and Japanese.
Summary written by a language model from the project’s public pages.
- Variable Bit Quantization
- Layer Sensitivity Analysis
- MLX Optimization
- Benchmark Comparison
- Memory Reduction
Tags
model-quantizationmlx-optimizationvariable-bit-widthllm-compressionsensitivity-aware
AI capabilities
Text
Inference: LocalWeights: Open
Built with & integrations
Hosting
Cloudflare
Runs on
CLI
Detected from
- Cloudflare
- cf-ray header · cf-cache-status header
Trust & compliance
License
Open Source
Verified signals
Indexing history
1What PulseGate has recorded for this listing
Frequently asked questions about JANGQ
- What does JANGQ do?
- Standard uniform quantization degrading performance in smaller LLMs by treating all layers with equal precision regardless of sensitivity. It is catalogued under AI & ML on PulseGate.
- Who should use JANGQ?
- JANGQ is an open-source project built for machine learning engineers and researchers.
- Is JANGQ free?
- Yes — JANGQ is open source under the Open Source license and free to use.
- What platforms does JANGQ run on?
- JANGQ runs on the command line.
- Is JANGQ still active?
- The GitHub repository shows 149 commits in the last 90 days.
- What projects are similar to JANGQ?
- Similar projects tracked by PulseGate include Qwen3.6 35B A3B CompQuant, Qwen3.6 35B A3B, and Qwen3.6 35B A3B.Qwen3.6 35B A3B CompQuantQwen3.6 35B A3BQwen3.6 35B A3B
- Who makes JANGQ?
- JANGQ is developed by JANG.
- How long has JANGQ been around?
- JANGQ first shipped in 2026.
Similar apps
Closest matches by what these projects do