Qwen2: the open model worth self-hosting for non-English filings

Most open-model releases this year are a single weight class chasing one leaderboard. Qwen2, from Alibaba, is a family: five sizes from 0.5B to 72B, including a 57B mixture-of-experts model, all using Group Query Attention, pretrained at 32K context and extended toward 128K. For a quant desk, the number that matters is not the English MMLU score, it is the Chinese and Asian-language coverage. That is the capability Western models handle worst and the one a global book needs most.

Qwen2-72B-Instruct posts the scores you want from a frontier-adjacent open model. MMLU is 82.3, GSM8K reaches 91.1, MATH is 59.7, and HumanEval is 86.0. The numbers that matter more for this use case are the multilingual ones: 78.0 on the multilingual MMLU and 86.6 on the multilingual grade-school math set, across 27 languages beyond English and Chinese.

Where it earns its place on a desk

The use case is not chat. It is batch document work over text that closed Western APIs read poorly. Parsing a Chinese exchange filing, a Japanese earnings release, or a Korean disclosure is exactly where a model trained heavily on those languages pulls ahead. A model that has seen a large volume of Chinese financial text will tokenize it more efficiently and misread its idioms less often than a model that treats it as a long-tail language. Running it on hardware you own means those filings never leave the building, which matters when the documents are price-sensitive.

The size ladder is the practical part. It is where the cost discipline lives. You do not put the 72B in a high-throughput scoring loop. You reach for a small Qwen2 to triage and tag thousands of documents cheaply, then escalate the handful that need real reasoning to the 72B. That tiering is how you keep a document pipeline affordable without sending everything to the most expensive call. The 0.5B and 1.5B models are small enough to run many in parallel on modest hardware, which is what makes the triage layer cheap.

Group Query Attention is the unglamorous reason the larger models are servable at all. It shrinks the memory cost of the attention cache at inference. A 72B model with a long context then stays within a realistic GPU budget. For a team that has to justify hardware, that efficiency is the difference between a model you can deploy and one you can only admire.

The licence, read carefully

One caveat that decides whether you can ship. The smaller models are Apache 2.0, but Qwen2-72B keeps the original Qianwen License, so read the terms before you build a commercial product on the flagship. The capability is open. The commercial freedom is conditional. The condition matters most for exactly the largest model you would want in production.

How I would use it

As the non-English workhorse in a document stack, behind a router. Score the easy cases with a small model, send the ambiguous ones to the 72B, and keep an English-strong model for the English filings where it still leads. Measure where the language gap actually costs you, because for an English-only corpus this is not the model to reach for. The lesson of Qwen2 is that open weights have reached the point where language coverage is the reason to pick one model over another for a given corpus, and for an Asian-market book that reason is decisive.

Qwen2’s edge for a quant desk is language coverage: it reads the Asian-market filings Western models fumble, on hardware you control, behind a router that keeps the cost down.