TLDRocket
Sign in

Qwen2.5-Math: The world's leading open-sourced mathematical LLMs

GitHub Pages

Qwen dropped a new open-source math LLM family, Qwen2.5-Math, and it's beating GPT-4o and Gemini's math specialist on tough benchmarks. The wild part: a 7B model now matches its old 72B version.

Based on reporting by GitHub Pages — read the original for the full story.

Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error

Alibaba's Qwen team just pushed out Qwen2.5-Math, and the numbers are the kind that make you double-check the source. The flagship 72B-Instruct model hits 92.9 on MATH when paired with tool-integrated reasoning and a reward model — beating GPT-4o and Google's math-specialized Gemini 1.5 Pro on the same test. That's not a marginal bump. It's the kind of gap that used to separate open and closed models by a full generation.

What's more interesting than the raw scores is how they got there. The team didn't just scale up more data — they used Qwen2-Math-72B-Instruct to synthesize fresh training data, then swapped in Qwen2.5's base models for better language and code understanding, and rebuilt the training corpus from 700 billion tokens to over a trillion. They also trained a dedicated 72B reward model to grade candidate solutions and used it both for rejection sampling during fine-tuning and for reinforcement learning with GRPO afterward. It's a pipeline that treats math ability less like a single training run and more like a refinery process, passing outputs back through the system to polish them.

The practical upshot is that model size stops mattering as much as it used to. Qwen2.5-Math-1.5B-Instruct, a model small enough to run on a laptop, now beats most 70B-class math models from a year ago. The 7B version actually surpasses the old Qwen2-Math-72B-Instruct on MATH, scoring 83.6 with chain-of-thought and 85.3 with tool-integrated reasoning, which lets the model write and execute Python instead of just talking through the steps. On AIME 2024, arguably the nastiest benchmark here, Claude 3 Opus, GPT-4 Turbo and Gemini 1.5 Pro each solve only 1 or 2 of 30 problems. Qwen2.5-Math-72B-Instruct solves 12 with TIR, and the 7B model, boosted by the reward model, gets up to 21.

None of this is a general-purpose chatbot upgrade — Qwen is explicit that this series is narrow, tuned for English and Chinese math problems and not much else. But narrow and open isn't a bad combination. If you need a model that can actually verify a derivative or solve a system of equations without hallucinating a wrong root, this is now sitting near the top of the pile, code and weights included, no API key required.

My take — AI-written commentary, not fact-checked reporting

I'll say what a lot of people are dancing around: a 1.5B open model outperforming last year's 70B giants is the actual headline here, not the leaderboard flex against GPT-4o. Efficiency gains like this are exactly why betting everything on ever-bigger closed models feels increasingly like a rich person's hobby rather than a technical necessity — and it's happening in the open, from a Chinese lab, while Western AI policy debates are still stuck arguing about whether open weights are dangerous.

Read more about this at: GitHub Pages

Related stories

The daily briefing

Every AI story that matters, in your inbox by 8am.

TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.