Qwen3.5 35B A3B (Alibaba Cloud)

Qwen3.5-35B-A3B from Alibaba Cloud / Qwen Team achieves 0.819 on the MMStar multimodal benchmark, placing third of 27 evaluated models. It sits behind Qwen3.6 Plus (0.833) and Qwen3.5-122B-A10B (0.829) while outperforming all listed Qwen3 VL instruct and thinking variants. The result demonstrates competitive vision-language reasoning at a 35B MoE size class. The MMStar leaderboard on llm-stats.com provides the broadest comparison set among supplied candidates, spanning 27 models including entries from ByteDance, LG AI Research, Liquid AI, DeepSeek, and Cohere. Qwen3.5-35B-A3B's third-place finish reinforces its standing as a strong mid-size multimodal model relative to both proprietary and open alternatives. Third-party leaderboard tracking as of mid-2026.

Qwen3.5-35B-A3B from Alibaba Cloud / Qwen Team achieves a score of 0.679 on the Hallusion Bench image-context reasoning benchmark, ranking third of 20 evaluated models. The benchmark tests 346 images with 1,129 questions assessing language hallucination and visual illusion in large vision-language models. The 35B-A3B variant trails Qwen3.5-27B at 0.700 and Qwen3.6-35B-A3B at 0.698. The Hallusion Bench leaderboard on llm-stats.com, published May 7, 2026, lists Qwen3.5-35B-A3B alongside Qwen3.5-27B, Qwen3.6-35B-A3B, and Qwen3.5-122B-A10B among the top four open proprietary performers. ByteDance's Seed 2.0 Mini and Seed 1.8 also feature in the top 11, placing the Qwen family competitively against other frontier model families on this hallucination-resistance evaluation.

Qwen3.5-35B-A3B from Alibaba Cloud / Qwen Team leads the AndroidWorld Success Rate agent benchmark with a score of 0.711, ranking first among 8 evaluated models. The 35B MoE variant outperforms larger siblings including Qwen3.5-122B-A10B at 0.664 and Qwen3.5-27B at 0.642 on this multimodal Android navigation benchmark. The result signals strong agentic capability for autonomous UI tasks. On the MMStar multimodal vision-language benchmark, Qwen3.5-35B-A3B ranks third of 27 models with a score of 0.819, trailing Qwen3.6 Plus (0.833) and Qwen3.5-122B-A10B (0.829). The 35B-A3B MoE architecture delivers competitive multimodal reasoning despite its mid-tier parameter count, exceeding all Qwen3 VL variants on this leaderboard. Third-party llm-stats.com tracking corroborates the model's multimodal benchmark performance.

Alibaba's Qwen team released the Qwen3.5-35B-A3B on February 25, 2026, as part of the four-model Qwen3.5 Medium series, available open-source under Apache 2.0 on Hugging Face and ModelScope for commercial use. The MoE model activates a fraction of its 35B total parameters, hybridizing Gated Delta Networks with sparse Mixture-of-Experts for efficiency. It supports agentic tool calling and is engineered to stay accurate under 4-bit weight and KV cache quantization. The Qwen3.5-35B-A3B reportedly exceeds one million tokens of context on consumer GPUs with 32GB of VRAM, enabling frontier-length context without server-grade hardware. According to Alibaba's own claims quoted in the coverage, it surpasses Qwen3-235B-A22B-2507, Qwen3-VL-235B-A22B, GPT-5-mini, and Claude Sonnet 4.5 on third-party benchmarks. A sibling Qwen3.5-Flash remains proprietary and is served only via the Alibaba Cloud Model Studio API.

GIGAZINE's machine-translated coverage confirms the February 25, 2026 release of the Qwen3.5 Medium Model Series by Alibaba's Tongyi Lab, listing Qwen3.5-35B-A3B alongside the 122B-A10B, 27B, and Flash variants. The article embeds the official @Alibaba_Qwen and @Ali_TongyiLab X announcements from February 24. It positions the release around the slogan "more intelligence, less compute," emphasizing the new model's efficiency claims. The GIGAZINE piece repeats Alibaba's assertion that Qwen3.5-35B-A3B surpasses both Qwen3-235B-A22B-2507 and Qwen3-VL-235B-A22B, framing the improvement as evidence that better architecture and data quality matter more than raw parameter count. The article is a secondary third-party source largely mirroring Qwen's own announcement text, adding little independent technical depth beyond corroborating the release event.

Qwen3.5-35B-A3B from Alibaba Cloud / Qwen Team scores 62.00% on the PMC-VQA biomedical multimodal question answering benchmark, placing third of three evaluated models. The benchmark is built on biomedical literature and medical figures. Evaluated September 27, 2026, the result trails Qwen3.5-122B-A10B at 63.30% and Qwen3.5-27B at 62.40%. The PMC-VQA leaderboard from llmboard.ai has a narrow comparison set of only three Qwen-family models, limiting comparative breadth. Within this small set, Qwen3.5-35B-A3B's 62.00% demonstrates medical multimodal VQA capability at the 35B-A3B MoE scale. The page provides third-party benchmark evidence with grade C reliability, supplementing broader multimodal evaluation signals from other leaderboards.

Qwen3.5-35B-A3B from Alibaba Cloud / Qwen Team scores 74.80% on the MVBench multimodal video understanding benchmark, placing fourth of 18 models at the 82.35th percentile. The evaluation was recorded on September 27, 2026. It ranks between Qwen3.6-27B at 75.50% and Qwen3.5-27B at 74.60% on temporal video reasoning tasks. The MVBench leaderboard from llmboard.ai covers 20 video task categories requiring temporal understanding beyond single-frame analysis, with 18 model participants. Qwen3.5-35B-A3B's score places it ahead of Qwen3.6-35B-A3B at 74.60% and several Qwen3 VL variants, confirming competitive video understanding at a mid-size MoE parameter class. Third-party benchmark evidence grade C.