Model

Qwen3.5-397B-A17B

OPEN

Alibaba · Qwen3.5 family · released Feb 16, 2026

Open-weight flagship of the Qwen3.5 generation; native vision-language model (not a bolt-on VL variant) built on Qwen3-Next-style hybrid linear-attention/MoE architecture. Native 262,144-token context, extensible to ~1.01M via YaRN. First confirmed model of the flagged 'missing generation'.

ReasoningCodingVisionFunction callingTool useAgentic

2182.0

Elo · rank #47

Parameters

397B

Active params

17B (MoE)

Context

256K tokens

Architecture

Hybrid-attention MoE (Qwen3-Next lineage): 60 layers, 15x(3x(Gated DeltaNet -> MoE) -> 1x(Gated Attention -> MoE)); 512 experts, 10 routed + 1 shared activated per token; hidden dim 4096; natively multimodal ("Causal Language Model with Vision Encoder" per HF card, not text-only)

License

Apache 2.0

Languages

201+

API price (in/out)

No hosted API

Modalities

text · vision · video

Benchmark results

Bar shows position within the tracked field; marker = field best

AI2DVision93.9%#11

best: Molmo 72B · 96.3%

AIMEMath91.3%#42

best: GPT-5.2 · 100.0%

Arena EloHuman preference1442#11

best: Claude Fable 5 · 1505

C-EvalKnowledge93.0%#2

best: Qwen3.6-Plus · 93.3%

CharXivVision80.8%#9

best: Muse Spark 1.1 (xhigh) · 88.4%

GDPval-AAAgents960#32

best: Claude Fable 5 · 1932

GPQA DiamondReasoning88.4%#34

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning28.7%#48

best: Claude Sonnet 5 · 57.4%

IFBenchReasoning76.5%#10

best: MiniMax M3 · 83.0%

IFEvalReasoning92.6%#8

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding83.6%#32

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MathVisionVision88.6%#3

best: Seed 2.1 Pro · 92.6%

MathVistaVision90.3%#2

best: Seed 2.1 Pro · 90.7%

MMBench (English)Vision93.7%#1

best: this model · 93.7%

MMLU-ProKnowledge87.8%#8

best: Claude Fable 5 · 91.5%

MMLU-ReduxKnowledge94.9%#2

best: Qwen3.7-Max · 95.0%

MMMLUKnowledge88.5%#12

best: Gemini 3.1 Pro · 92.6%

MMMU-ProVision79.0%#11

best: Claude Opus 4.7 · 85.5%

MMMUVision85.0%#8

best: Claude Fable 5 · 89.3%

OCRBenchVision93#14

best: InternVL3-78B · 906

OSWorld-VerifiedAgents62.2%#19

best: Claude Fable 5 · 85.0%

SuperGPQAReasoning70.4%#5

best: Qwen3.7-Max · 73.6%

SWE-bench VerifiedCoding76.4%#37

best: Claude Fable 5 · 95.0%

Terminal-Bench 2.0Coding52.5%#43

best: GPT-5.6 · 88.8%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Qwen3.5 family

Elo progression across releases

Qwen3.5-397B-A17BFeb 20262182.0 Qwen3.5-PlusFeb 20261882.6 Qwen3.5-122B-A10BFeb 20262021.4 Qwen3.5-Omni-FlashMar 20261471.5 Qwen3.5-Omni-PlusMar 20261834.0

Resources

HFQwen/Qwen3.5-397B-A17B GHQwenLM/Qwen3.6 DOCShttps://huggingface.co/Qwen/Qwen3.5-397B-A17B APIAvailable via Alibaba API

API price weights · each benchmark row carries its own source badge (see methodology)