Qwen3-Next-80B-A3B-Thinking is a reasoning-first chat model in the Qwen3-Next line that outputs structured “thinking” traces by default. It’s designed for hard multi-step problems; math proofs, code synthesis/debugging, logic, and agentic planning, and reports strong results across knowledge, reasoning, coding, alignment, and multilingual evaluations. Compared with prior Qwen3 variants, it emphasizes stability under long chains of thought and efficient scaling during inference, and it is tuned to follow complex instructions while reducing repetitive or off-task behavior. The model is suitable for agent frameworks and tool use (function calling), retrieval-heavy workflows, and standardized benchmarking where step-by-step solutions are required. It supports long, detailed completions and leverages throughput-oriented techniques (e.g., multi-token prediction) for faster generation. Note that it operates in thinking-only mode.
Evaluations
10
across 10 benchmarks
Latency
1m 14s
Context length
262k
tokens
Cost
$0.15 · $1.20
input · output per 1M tokens
Key takeaways
The Qwen3 Next 80B A3B Thinking is a reasoning-first chat model best suited for complex, multi-step problems like math proofs, code synthesis/debugging, and agentic planning, operating in a "thinking-only" mode.
It employs a Mixture of Experts (MoE) architecture, emphasizing stable, long chains of thought and efficient scaling with techniques like multi-token prediction for faster generation and detailed completions.
The model demonstrates strong logical and mathematical reasoning, achieving 86.6% accuracy on the AGI-NExT Benchmark and 86.7% on the AIME 2024 benchmark, particularly excelling in mathematical problem-solving and converting word problems into equations.
Despite its strengths, it struggles with precision, factual exactitude, and nuanced interpretation in specialized or ambiguous domains, often providing confidently incorrect answers (e.g., 2.22% accuracy on "Humanity's Last Exam").
Common failure modes include factual inaccuracies, misinterpretation of subtle language, calculation errors in multi-step problems, and confidently incorrect reasoning, especially in highly specialized or nuanced fields like combinatorics and certain general knowledge tasks.
