Qwen3.6 35B-A3B

A practical Medium tutorial demonstrates running Qwen3.6-35B-A3B locally via llama.cpp on consumer hardware with just 6 GB of VRAM and 32 GB of RAM, achieving approximately 30 tokens per second. The article, published May 10, 2026, highlights that recent updates to llama.cpp combined with the model's sparse 35B/3B MoE The piece frames this as a significant improvement over prior setups where larger models required substantial VRAM and were impractical locally, while smaller models were limited to 32K–128K contexts. By leveraging Qwen3.6-35B-A3B's sparsity and recent llama.cpp optimizations, the author shows a path to running frontie

BenchLM aggregates verified and provisional evidence for Qwen3.6-35B-A3B, recording category scores of 40.6 in agentic, 45.3 in coding, and 50.0 in multimodal. Reasoning, multilingual, and instruction following categories are listed as not measured. The dashboard ranks it 39th in multimodal and grounded workflows. The aggregator reports a 262K token context, median speed of 115 tokens per second, and a first-token latency of 48.94 seconds. Capability is scored at 43.6 out of 100 against a field median of 56.6, with no first-party API token rate available.

Developer-reported benchmarks for Qwen3.6-35B-A3B list SWE-bench Verified at 73.4 percent resolved, SWE-bench Pro at 49.5 percent pass@1, Terminal-Bench 2.0 at 51.5 percent tasks resolved, LiveCodeBench at 80.4 percent pass@1, and MMLU-Pro at 85.2 percent accuracy. The model card attributes development to Alibaba's Qwen team. The model card reports a 262K token context window and an April 2026 release for the 35B/3B active MoE variant. A hardware fitting table shows FP16 weights near 70GB while IQ3 quantizations compress to about 14GB, enabling runs on consumer GPUs like the RTX 3060 12GB with partial CPU offload.

Alibaba's Tongyi Lab released Qwen3.6-35B-A3B as an open model on April 15, 2026, under the Apache 2.0 license. The launch followed the April 2, 2026 announcement of Qwen3.6-Plus and positions the variant for agentic coding workflows. Tongyi Lab markets it as stronger than Google's Gemma 4 open model suite. Qwen3.6-35B-A3B is a sparse mixture-of-experts model with 35 billion total parameters and only 3 billion active parameters, designed for efficient local inference. Simon Willison demonstrated the model drawing a pelican on a laptop that beat Claude Opus 4.7 in his qualitative test. Official details are documented on the qwen.ai blog linked from the coverage.

Alibaba's Qwen team open-sourced Qwen3.6-35B-A3B in April 2026 as the first open-weight variant of the Qwen3.6 generation, releasing weights under Apache 2.0 on Hugging Face alongside the proprietary Qwen3.6-Plus API model. The 35-billion-parameter sparse Mixture-of-Experts activates only 3B parameters per token and is Architecturally, Qwen3.6-35B-A3B uses 256 experts with 8 routed plus 1 shared expert activating per token in a 40-layer stack that repeats a 3:1 block of Gated DeltaNet linear-attention layers followed by a Gated Attention layer, each paired with an MoE feed-forward block, with a hidden dimension of 2048 and expert int