Qwen3.6 35B-A3B

A Medium article by Minyang Chen describes running Qwen3.6-35B-A3B locally via llama.cpp, claiming speeds of around 30 tokens per second on a 6GB VRAM consumer machine. The piece notes the model was released by Alibaba in April 2026 as a sparse Mixture-of-Experts design with 35 billion total parameters activating only 3 billion per token, targeting coding and reasoning workloads. Context window support up to 256K tokens is highlighted for low-end hardware deployments. The article frames the Qwen3.6 generation as a notable efficiency step, allowing larger sparse models to run on older CPUs and limited VRAM at speeds comparable to much smaller dense models. It positions the 35B-A3B variant as bridging the gap between resource-heavy frontier models and deployable local inference for developers and hobbyists using llama.cpp tooling.

A Towards AI article by Mustafa Genc discusses Qwen3.6-35B-A3B as the first open-weight model in the Qwen3.6 generation, released by Alibaba in April 2026. The piece addresses the "goldfish-brain" problem in agentic LLM deployments, where models lose their reasoning chain across turns, and positions the upgrade as preserving thinking state in multi-turn agent workflows. The article notes Qwen3.6-35B-A3B shares its hybrid sparse MoE skeleton with the prior Qwen3.5-35B-A3B, using Gated DeltaNet linear attention combined with standard gated attention in a 3:1 hybrid ratio across 256 experts. It frames the 3.5-to-3.6 transition as focused on agentic reasoning continuity rather than scaling or architectural redesign, emphasizing practical impact for AI agent developers.

BenchLM tracks Qwen3.6-35B-A3B with a composite capability score of 44.7 out of 100 (rank 115 of 211) as of September 30, 2026, and reports a generation speed of 126 tokens per second with a 262K-token context window. The model ranks 37th of 49 in the multimodal category at a score of 51.0, reflecting particular strength in screenshots, documents, charts, and grounded multimodal workflows. Category-level verified evidence includes agentic at 28.9 (rank 83 of 119, 11 benchmarks), coding at 35.6 (rank 79 of 144, 6 benchmarks), and knowledge at 42.0 (rank 94 of 170, 5 benchmarks), while reasoning, multilingual, and instruction-following categories show zero measured benchmarks. No first-party hosted API rate is published for direct cost comparison.

MillstoneAI's on-prem inference benchmark evaluates Qwen3.6-35B-A3B in FP8 precision on a single MI300X with 192GB VRAM using vLLM, without prompt caching or speculative decoding. Across 1K to 128K context lengths and 1 to 3 concurrent requests, the configuration maintained a 100.0% success rate over the test scenarios. Peak system throughput reached 255.6 tokens per second at 3 concurrent requests with 1K context, supporting 3 concurrent chatbot users at 32K context. Single-request generation scaled from 113.9 tok/s at 1K context down to 7.3 tok/s at 128K context, with time-to-first-token rising from 116 ms to 27.1 s across the same range.

llm-stats ranks Qwen3.6-35B-A3B at position 113 overall, with strong top-decile placement in healthcare and legal categories (18 of 248 and 21 of 213) but below median in tool calling and coding. The tracker lists a blended price of $0.14 per million tokens against an LLM Stats Score of 32.4, with logged benchmark sources from qwen.ai. Verified benchmark rows from the scorecard include 0.93 on MMLU-Redux (rank 11), 0.93 on MMBench-V1.1 English dev (rank 1), 0.93 on AI2D (rank 8), and a tracked AIME 2026 entry, providing independent confirmation of the model's multimodal and reasoning performance.

On April 2, 2026, Alibaba's Qwen team open-sourced Qwen3.6-35B-A3B under Apache 2.0, the first open-weight variant of the Qwen3.6 generation. The 256-expert sparse Mixture-of-Experts model activates 8 routed plus 1 shared expert per token for 3B active parameters, using a hybrid stack of three Gated DeltaNet linear-attention layers followed by one Gated Attention layer. Qwen3.6-35B-A3B supports text, image, and video via a built-in vision encoder with a native 262,144-token context extensible to roughly 1.01M with YaRN, and runs on Hugging Face Transformers, vLLM 0.19+, SGLang 0.5.10+, and KTransformers. It reports frontier-level scores including 73.4 on SWE-bench Verified and 92.7 on AIME 2026, and introduces thinking preservation so multi-step agentic workflows retain reasoning across tool calls.