Qwen3-Coder-30B-A3B-Instruct-GGUF

Qwen3-Coder-30B-A3B-Instruct-GGUF is a quantized (GGUF) distribution of Alibaba's Qwen3-Coder-30B-A3B-Instruct, an open-source Mixture-of-Experts (MoE) coding model released under Apache 2.0 on July 31, 2025. It has 30.5 billion total parameters with 3.3 billion activated per forward pass across 128 experts (8 active), using 48 layers with Grouped-Query Attention. The model is optimized for agentic coding, repository-scale understanding, function calling, and tool use via platforms such as Qwen Code and Cline.

On benchmarks, the full-precision instruct model scores strongly on coding tasks: HumanEval ~92.7% (independent local eval) and GSM8K ~90.67% strict-match (vLLM lm-eval, BF16). On general knowledge, MMLU reaches ~84.7% (academic BF16 baseline on the A3B instruct family). Official agentic evaluations report 51.6% on SWE-bench Verified with OpenHands scaffolding. The model natively supports 262,144 tokens of context (extendable to 1M via YaRN), making it well-suited for whole-repo coding workflows.

The GGUF variant (commonly from unsloth or community quantizers) enables local deployment via llama.cpp, Ollama, LM Studio, and KTransformers. At Q4_K_M quantization—the community default—weights require ~18–20 GB VRAM, fitting single 24 GB consumer GPUs. Recommended sampling: temperature 0.7, top_p 0.8, top_k 20. For OOM issues, reduce context to 32,768 tokens.