Unique Research · Models · November 28, 2024
Alibaba’s Qwen team released QwQ-32B-Preview — among the few models pitched as rivaling OpenAI’s GPT reasoning line, and the first in that class available for download under a permissive license.
32.5B parameters; prompts up to ~32,000 words. On some benchmarks it beats OpenAI’s two reasoners so far — o1-preview and o1-mini (CN text also references GPT-3.5 Turbo / GPT-3 naming; treat as o1-preview / o1-mini). Parameter count roughly tracks problem-solving capacity; OpenAI does not disclose its counts.
Alibaba internal tests: QwQ-32B-Preview tops OpenAI GPT reasoners on AIME and MATH. Strength: “reasoning” for logic puzzles and hard math. Limits (company blog): may switch languages unexpectedly, loop, and struggle on “common-sense reasoning.”
Like peer reasoners, it self-fact-checks — fewer classic traps, longer time to answer. It plans tasks and runs steps toward an answer, GPT-style.
Released under Apache 2.0 for commercial use — but only some components shipped, so you can’t fully reproduce or inspect internals. “Openness” is a spectrum; QwQ sits between API-only and full weights+data openness.
As scaling-law returns draw scrutiny — reports that OpenAI, Google, Anthropic frontier gains have slowed — reasoners and test-time / inference compute (extra processing time at answer time) get more attention. Beyond OpenAI and Chinese labs, The Information: Google grew its reasoning-model team to ~200 people and added substantial compute.
— END —