Club-3090 is a community-driven repository that provides working configurations for running modern large language models on RTX 3090 GPUs. Created by developer 'noonghunna' and released under Apache 2.0 license, the project has gained 202 stars on GitHub since launching in late April 2026.
Production-Ready Qwen3.6-27B on 24GB Consumer Hardware
The project consolidates configurations, patches, and benchmarks for serving state-of-the-art LLMs locally on consumer-grade RTX 3090 hardware with 24GB VRAM. The current focus is Qwen3.6-27B, which the project marks as production-ready, with additional models like Qwen3.5-27B and GLM-4.6 following the same pattern.
Club-3090 provides drop-in OpenAI-compatible API access, enabling developers to run powerful models on hardware many already own without requiring enterprise-grade infrastructure.
Two Complementary Approaches for Different Priorities
The repository offers two deployment strategies:
vLLM Dual-Card Route emphasizes throughput with up to 127 tokens per second for code generation using DFlash or 4 concurrent streams at 262K context length. This approach supports vision capabilities, tool use, and multi-token prediction with streaming.
Llama.cpp Single-Card Route prioritizes stability, delivering full 262K context on one RTX 3090 without prefill cliffs. While slower at approximately 21 tokens per second, this route handles 25K-token tool returns and passes 90K needle ladder stress tests with greater robustness.
Comprehensive Benchmarking and Community Resources
The project includes standardized benchmark protocols involving 3 warm-up runs plus 5 measured runs of canonical narrative and code prompts on RTX 3090 at 230W. Testing uses vLLM nightly builds and llama.cpp mainline to ensure reproducible results.
Technical requirements include 1-2 RTX 3090s (larger Ampere/Ada cards also compatible), Linux (Ubuntu 22.04+), Docker with NVIDIA Container Toolkit, NVIDIA driver 580.x+, and approximately 30GB storage per model. The vLLM route requires CUDA and Linux exclusively.
Democratizing Access to State-of-the-Art Models
Club-3090 removes trial-and-error barriers for local LLM deployment by providing tested configurations and interactive setup wizards. The project credits contributors across the local-LLM community and acknowledges upstream dependencies including the Qwen team, Genesis patches, and Intel AutoRound.
Developers can access canonical benchmark scripts for performance validation and comprehensive documentation covering hardware setup, engine selection, and model internals. The repository invites community participation through issue reporting and pull requests for new model variants.
Key Takeaways
- Club-3090 provides production-ready configurations for running Qwen3.6-27B on RTX 3090 GPUs with 24GB VRAM, gaining 202 GitHub stars since April 2026
- Two deployment approaches offer different trade-offs: vLLM dual-card for throughput (127 TPS) and llama.cpp single-card for stability (21 TPS with 262K context)
- The project delivers OpenAI-compatible API access, enabling drop-in replacement for cloud LLM services
- Standardized benchmarking protocols ensure reproducible performance validation across community configurations
- Apache 2.0 licensing and community-driven development democratize access to state-of-the-art models on consumer hardware