Qwen3.5 397B-A17B

Artificial Analysis ranks Qwen3.5-397B-A17B third among open-weights models on the Intelligence Index at 45, a 16-point gain over Qwen3-235B-A22B-2507, behind GLM-5 and Kimi K2.5. It is the first Qwen open-weights model with native vision input, unifying the previous separate text-only and Qwen3-VL lines. Key improvements include GDPval-AA ELO of 1,221 (+361 over Qwen3 235B), +27 points on TerminalBench Hard, +12 on HLE, and +28 on IFBench. The model supports both reasoning and non-reasoning modes in a single checkpoint, a reversal from Qwen3's split instruct/thinking variants, though hallucination rate remains high at 88%.

Third-party evaluator llm-stats.com profiles Qwen3.5-397B-A17B with a composite LLM Stats Score of 13.9 and a global rank of 70 across tracked benchmarks. The model achieves tier-S standing in Finance (rank 2 of 228), Legal (4 of 213), and Healthcare (4 of 248), placing it in the top 2% in those categories. These category-level rankings directly reflect the model's evaluation performance across knowledge-intensive domains. The profile also shows tier-A performance in Chat (3 of 133) and Audio (5 of 81), with tier-B scores in Long Context, Math, Reasoning, Tool Calling, Coding, and Vision. A conversation-depth analysis reveals a 1.9-point decline in score after turn 30, while the Quality Tracker marks the model as stable at +0.98σ over 310 votes. Dataset-level scores such as MMLU-Redux (0.95) and HMMT 2025 are attributed to qwenlm.github.io, though the underlying Qwen sources are not independently supplied.

BenchmarkList tracks 132 benchmark results for Qwen3.5 397B A17B, comparing it against Claude Opus 5.5, Claude Fable 5.1, MiMo V2.6 Pro, Kimi K3, and Qwen3.8 variants. The model ranks competitively across agentic, coding, and reasoning evaluations, with τ²-bench at 87.9% and Vals.ai Financial Agent 1.1 at 61.3%. On ClawProBench the model scores 64.18 overall, with 71.6% tool use, 77.8% planning, 72.4% constraints handling, and 74.5% error recovery. Tau2-Bench Telecom shows a 95.6% success rate, indicating strong agentic performance across multi-step tool-use and conversational AI scenarios.

evals.report tracks 22 Qwen3.5-397B-A17B benchmark scores from its February 16, 2026 release, covering SWE-bench Verified (76.4%), GPQA Diamond (88.4%), Humanity's Last Exam (28.7%), Berkeley Function Calling (72.9%), and MMMU-Pro (79.0%), with each result tagged by source verification status and date. Additional results include τ²-bench Telecom 95.6% pass rate, AIME 2026 at 93.33%, GDPval ELO 1220, Epoch Capabilities Index 146.1, and Terminal-Bench entries. Scores are reported individually rather than combined into a single ranking, giving developers traceable evidence across coding, reasoning, tool use, and multimodal categories.

apxml documents Qwen3.5-397B-A17B as a 397B-parameter MoE model with 17B active parameters across 512 experts, 262K context, and multimodal capability, released February 23, 2026 under Apache 2.0. It uses grouped-query attention with 32 Q heads and 2 KV heads, RoPE positional encoding, and SwiGLU activations over 60 layers. Benchmarks include MMLU Pro 87.8%, GPQA Diamond 88.4%, SWE-bench Verified 80.0%, and Terminal-Bench 2.0 54.0%. VRAM requirements are roughly 742 GB at 1K context in FP16, with recommended configurations including 9x RTX 6000 Blackwell, 3x Blackwell B300, or 2x M3 Ultra systems.

Alibaba announced Qwen3.5 on February 16, 2026, open-sourcing Qwen3.5-397B-A17B, also called Qwen3.5-Plus, as a natively multimodal foundation model designed for efficient inference. The model combines linear attention with sparse mixture-of-experts, supports 201 languages, and was trained on trillions of vision-language tokens across text, images, video, STEM, and reasoning data. The release targets reasoning, coding, agentic workflows, and image/video comprehension, aiming to lower deployment cost for agentic AI. Alibaba frames it as advancing scalable multimodal agents, unifying text, image, and video understanding in a single model that can process multimodal input and generate text output.

Alibaba's Qwen3.5-397B-A17B is the first open-weights release in the Qwen3.5 series, a native vision-language model with 397B total and 17B activated parameters, released in mid-February 2026 under Apache 2.0. It targets reasoning, coding, agent tasks, and multimodal understanding, unifying thinking and non-thinking behavior in a single checkpoint that drops Qwen3's soft /think switches. InferenceX reports an Artificial Analysis Intelligence Index score of 45, ranking third among open-weights models, with a GDPval-AA agentic ELO of 1,221 versus 860 for Qwen3 235B. The model uses multi-token prediction and is compatible with Transformers, vLLM, SGLang, and KTransformers, with a hosted Qwen3.5-Plus counterpart offering 1M-token context.