Qwen3-Coder 30B-A3B Instruct

An independent benchmark report on Level1Techs evaluates qwen3-coder:30b (Q4_K_M via Ollama) against the Beneficial AI Foundation's vericoding set, a SPARK/Ada formal verification benchmark in which a theorem prover must discharge every proof obligation a specification generates rather than relying on test cases. On 26 The steeper-than-frontier collapse on harder specs is presented as the key caveat for developers considering the model for proof-bearing or specification-heavy work. Setup details — Q4_K_M quantization, Ollama runtime, single box — are fully disclosed and reproducible, but the results are limited to the SPARK/Ada track

LLM Explorer's model page explicitly names "Qwen3 Coder 30B A3B Instruct" and mirrors the Hugging Face repository metadata: 30B parameter model, 61.1 GB required VRAM, 262,144-token (256K) context length, Apache 2.0 license, Qwen3MoeForCausalLM architecture, bfloat16 weights, Qwen2Tokenizer with a 151,936-token vocabul The page also lists the available sharded weight files (16 shards of 4.0 GB each plus a 1.1 GB final shard, totaling around 61.1 GB), shows a Transformers version requirement of 4.52.3, and presents community-quantized alternatives such as AWQ 4-bit, GPTQ 4-bit, 8-bit, and FP8 builds with their download counts and VRAM

Qwen3-Coder-30B-A3B-Instruct in FP8 precision is documented as a 30.5 billion-parameter Mixture-of-Experts (MoE) model featuring 128 experts with 8 active per forward pass and 3.3 billion parameters activated at inference time. The model is built on the Qwen3 architecture with 48 layers and grouped-query attention (32 The variant natively supports a 256,000-token context length, extendable to one million tokens via YaRN, making it well-suited for repository-scale code understanding, agentic coding tasks, browser-use scenarios, and structured tool calling through OpenAI-compatible formats. Millstone's page also reports independent in