Based on the new Qwen3‑Next architecture (hybrid attention, highly sparse MoE, training‑stability optimizations, and multi‑token prediction), the Qwen3‑Next‑80B‑A3B‑Instruct model delivers extreme efficiency with only 3B active parameters per pass. It performs comparably to Qwen3‑235B‑A22B‑Instruct‑2507 and shows clear advantages on ultra‑long context tasks (up to 256K tokens).
Added Sep 11, 2025
Model weightsContext Window
256.0K
Max Output
262.1K
Avg output tokens (7d)
48 tokens
Input Price (Auto)
$0.090/1M
Output Price (Auto)
$1.10/1M
Cache Read (Auto)
$0.045/1M
Capabilities
Benchmarks
Benchmarks
Performance metrics and benchmarks
Sourced from Artificial Analysis.
Intelligence Index
9.6
Agentic work
T²-Bench Telecom (legacy)
Legacy fallback · Conversational AI agents in dual-control scenarios
21.6%
Better than 22% of models compared
Document reasoning
AA-LCR v1.1
Long context reasoning with updated grading
52.7%
Better than 48% of models compared
Reasoning
HLE
Humanity's Last Exam
7.6%
Better than 47% of models compared
IFBench
Instruction-following benchmark
39.7%
Better than 38% of models compared
CritPt
Research-level physics reasoning
0.0%
Coding
Terminal-Bench Hard (legacy)
Legacy fallback · Agentic coding and terminal use
7.6%
Better than 40% of models compared
LiveCodeBench
Contamination-free coding benchmark
68.4%
Better than 76% of models compared
Math
AIME 2025
American Invitational Mathematics Examination 2025
66.3%
Better than 61% of models compared
Knowledge
MMLU-Pro
Professional and academic subject knowledge
81.9%
Better than 77% of models compared
AA-Omniscience Accuracy
Proportion of correctly answered questions
17.2%
AA-Omniscience Hallucination Rate
Rate of incorrect answers among non-correct responses
92.7%
Legacy benchmarks
GPQA Diamond (legacy)
Graduate-level scientific reasoning
73.8%
Better than 54% of models compared
AA-LCR (unversioned / legacy)
Long context reasoning evaluation
52.7%
Better than 48% of models compared
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 Next 80B A3B (Instruct) with similar models from the same provider or model family.
Qwen3 Next 80B A3B (Thinking)
qwen/qwen3-next-80b-a3b-thinkingQwen3 Next 80B A3B (Thinking)
Qwen3.6 35B A3B Thinking
qwen/qwen3.6-35b-a3b:thinkingQwen3.6 35B A3B is a native vision-language MoE model with hybrid attention. Compared to Qwen3.5 35B A3B, Alibaba reports stronger agentic coding, mathematical and code reasoning, and better spatial understanding (including object localization and detection).
Qwen3.6 35B A3B
qwen/qwen3.6-35b-a3bQwen3.6 35B A3B is a native vision-language MoE model with hybrid attention. Compared to Qwen3.5 35B A3B, Alibaba reports stronger agentic coding, mathematical and code reasoning, and better spatial understanding (including object localization and detection).
Qwen3 Coder Next
qwen/qwen3-coder-nextQwen3 Coder Next is an open-weight coding model built on Qwen3-Next-80B-A3B-Base (hybrid attention + MoE). It is agentically trained at scale on executable tasks and environment interaction, delivering strong coding and tool-use performance at lower inference cost. Native 256K context.
Qwen 3 235b A22B 2507
qwen/qwen3-235b-a22b-instruct-2507Qwen 3 235b A22B Instruct 2507 the updated version of Qwen3 235B A22B, with significant improvements in performance. This model is non-thinking.
Qwen3 30B A3B
qwen/qwen3-30b-a3bQwen 3 30b A3B is a 30b model with 3 billion active parameters per pass. Supports switching between thinking and non thinking: trigger thinking with /think and /no_think anywhere in a prompt or system message to toggle chain-of-thought reasoning.