Qwen 3.5's open-source 397B MoE model (17B active params) with hybrid linear attention and extended reasoning. Supports text, image, and video input with a 256K context window.
Added Feb 16, 2026
Model weightsContext Window
258.0K
Max Output
65.5K
Avg output tokens (7d)
3.6K tokens
Input Price (Auto)
$0.39/1M
Output Price (Auto)
$2.34/1M
Cache Read (Auto)
$0.20/1M
Capabilities
Benchmarks
Benchmarks
Performance metrics and benchmarks
Sourced from Artificial Analysis.
Intelligence Index
19.1
Coding Index
48.2
Agentic Index
10.6
Agentic work
AutomationBench-AA
Workflow automation with guardrail penalties
6.9%
Better than 38% of models compared
AA-Briefcase
Agentic knowledge work (Elo)
563 Elo
Better than 29% of models compared
GDPval-AA v2
Economically valuable tasks (Elo)
905 Elo
Better than 43% of models compared
Document reasoning
GDP.pdf
Professional PDF reasoning: all-pass rate
12.2%
Better than 54% of models compared
AA-LCR v1.1
Long context reasoning with updated grading
77.3%
Better than 81% of models compared
Reasoning
HLE
Humanity's Last Exam
29.0%
Better than 80% of models compared
IFBench
Instruction-following benchmark
78.8%
Better than 97% of models compared
CritPt
Research-level physics reasoning
1.7%
Coding
Terminal-Bench v4.0
Practical coding and terminal tasks
0.0%
Better than 18% of models compared
SciCode
Python programming for scientific computing
44.8%
Better than 41% of models compared
Knowledge
AA-Omniscience Accuracy
Proportion of correctly answered questions
30.8%
AA-Omniscience Hallucination Rate
Rate of incorrect answers among non-correct responses
88.9%
Legacy benchmarks
GPQA Diamond (legacy)
Graduate-level scientific reasoning
89.3%
Better than 87% of models compared
Terminal-Bench Hard (legacy)
Agentic coding and terminal use
40.9%
Better than 87% of models compared
T²-Bench Telecom (legacy)
Conversational AI agents in dual-control scenarios
95.6%
Better than 95% of models compared
AA-LCR (unversioned / legacy)
Long context reasoning evaluation
77.3%
Better than 81% of models compared
GDPval-AA (unversioned / legacy)
Economically valuable tasks
20.3%
Last updated Sep 11, 2026
Artificial AnalysisProviders
Choose explicit providers for this model. Auto routing remains available as the default option.
Loading provider options…
Related text models
Compare Qwen3.5 397B A17B Thinking with similar models from the same provider or model family.
Qwen3.5 397B A17B
qwen/qwen3.5-397b-a17bQwen 3.5's open-source 397B MoE model (17B active params) with hybrid linear attention. Supports text, image, and video input with a 256K context window.
Qwen 3.8 27B Queen
qwen/qwen3.8-27b-queenQwen 3.8 27B Queen is an open-weight roleplay finetune with image understanding, tool calling, optional reasoning, and a 262,144-token context window.
Qwen 3.8 27B Fable
qwen/qwen3.8-27b-fableQwen 3.8 27B Fable is an open-weight multimodal creative finetune for expressive dialogue, long-form storytelling, character work, and roleplay.
Qwen 3.8 27B Obliterated
qwen/qwen3.8-27b-obliteratedQwen 3.8 27B Obliterated is an open-weight multimodal model LoRA-tuned for fewer refusals across chat, coding, reasoning, tool use, and long-context work.
Qwen 3.8 27B Obliterated Thinking
qwen/qwen3.8-27b-obliterated:thinkingQwen 3.8 27B Obliterated with thinking enabled for more deliberate creative work, coding, multimodal analysis, tool use, and long-context problem solving.
Qwen 3.8 27B Uncensored
qwen/qwen3.8-27b-uncensoredQwen 3.8 27B Uncensored is an NVFP4 open-weight multimodal model LoRA-tuned for fewer refusals across chat, coding, reasoning, tool use, and long-context work.