Qwen3-Coder 30B-A3B Instruct

LLM Explorer lists Qwen3 Coder 30B A3B Instruct as an open-source model maintained by Qwen, with a reported VRAM footprint of 61.1 GB at full precision and a 256K context length under the Apache-2.0 license. The page records the Hugging Face repository as the source of truth, names the architecture as Qwen3MoeForCausal The model is sharded across sixteen roughly 4.0 GB safetensors files plus a final 1.1 GB shard, stored in bfloat16, which is consistent with self-hosted deployment of the full-precision variant. LLM Explorer also catalogs community quantized derivatives of Qwen3 Coder 30B A3B Instruct, including AWQ 4-bit, GPTQ 4-bit,

Millstone AI published an inference benchmark page for the FP8-quantized variant of Qwen3-Coder-30B-A3B-Instruct, confirming it is a 30.5B parameter MoE model with 128 experts and 8 active per forward pass, activating only 3.3B parameters at inference time. The page documents 48 layers with GQA (32 query heads, 4 KV he The benchmark reports measured throughput and capacity across hardware: 334 tok/s on a single RTX Pro 6000 Blackwell 96GB, 584 tok/s on a single H100 SXM 80GB, and 600 tok/s on a single H200 SXM 141GB, with concurrency and context-length ranges detailed per device. Capacity planning tables list concurrent request limit