Qwen2.5 72B Instruct in 2026: When a 2024 Model Still Fits

Qwen2.5 72B Instruct was released on 19 September 2024. In model terms that is several generations ago, and the Qwen team has shipped Qwen3, Qwen3.5, Qwen3.6 and more since. Yet it is still listed by most inference providers and still downloaded. The question worth answering is not whether it is the best model, which it plainly is not, but why anyone still picks it and whether you should.

What the model is

It is a dense transformer with 72.7B parameters, 70.0B of them outside the embeddings, across 80 layers. Attention is grouped-query with 64 query heads and 8 key-value heads. The model card lists RoPE, SwiGLU and RMSNorm, the standard recipe of its time, with BF16 weights.

Dense matters here. Every token touches all 72.7B parameters, which makes it more expensive per token to serve than a modern mixture-of-experts model of similar quality, and much more expensive than one with a few billion active parameters. Per-token compute, not total size, is what sets the serving bill.

The Qwen2.5 release post says the family was pretrained on up to 18 trillion tokens and claims improvements over Qwen2 in knowledge, coding, maths, instruction following, long-form generation and JSON output, with support for over 29 languages. At release the team positioned the 72B base model as competitive with Llama 3 405B, a much larger model.

The context window is smaller than it looks

This is the detail that causes the most confusion. The release post advertises 128K tokens of input and up to 8K tokens of generation. The model card is more precise: the shipped configuration is 32,768 tokens, and reaching 131,072 requires enabling YaRN rope scaling with a factor of 4.

The card also warns that static YaRN scaling can hurt quality on short inputs, and recommends adding it only when long context is actually needed. So in practice there are two models: a 32K one that behaves as tested, and a 128K one that trades some short-prompt quality for reach.

Check which one a hosted endpoint serves before you plan around it. OpenRouter lists it with a 32,768-token context. If you need more than that, self-host with YaRN enabled, or more sensibly, use a newer model with a native long window.

The licence is not Apache 2.0

Most of the Qwen2.5 family is Apache 2.0. The release post is explicit that the 3B and 72B models are the exceptions. The 72B ships under the Qwen Licence Agreement, and two clauses matter commercially:

• The 100 million user threshold. If your product or service has more than 100 million monthly active users, you must request a licence from Alibaba Cloud before using it commercially.
• Attribution on derivatives. If you use the model or its outputs to train or fine-tune a model that you distribute or make available, you must display "Built with Qwen" or "Improved using Qwen" in the product documentation.

Neither is a problem for most teams, but the second one is easy to miss in a distillation pipeline. The licence also sets Chinese law and Hangzhou courts as governing. Open weights vs open source covers why these custom terms deserve a read before you ship.

Why people still use it

There are three honest reasons, and none of them is raw capability.

1. It is a known quantity. Two years of production use means its failure modes, prompt quirks and quantisation behaviour are well documented. For a regulated workflow that was validated against it, re-validating a new model is real work.
2. Existing fine-tunes. A large ecosystem of fine-tuned derivatives and adapters was built on it. If your adapter was trained on this base, switching base means retraining.
3. Price on some aggregators. OpenRouter listed it at $0.36 per million input tokens and $0.40 per million output at the time of writing. The near-flat input and output pricing suits output-heavy work such as long-form generation.

If none of those applies to you, a newer Qwen model is almost certainly the better default. The Qwen3 and later generations offer hybrid reasoning, much longer native context and, in the mixture-of-experts variants, far lower per-token compute.

What to compare it against

Within its own family, the Qwen2.5 7B Instruct guide covers the small sibling, a tenth of the size and fine for classification or extraction that does not need the 72B's depth. For code, the Qwen2.5 Coder 32B Instruct was the specialised option of the same generation.

Against newer models, test rather than assume. Take twenty prompts from your real workload, run them through 72B and through one current Qwen model at a similar or lower price, and score them blind. If the old model does not lose on your tasks, its stability may be worth keeping. If it does, the migration is usually a model id change plus a prompt review.

A decision rule

Keep Qwen2.5 72B Instruct if you have a validated workflow, an adapter trained on it, or an output-heavy job where its pricing is genuinely lower. Do not start new work on it. And if you keep it, plan the exit: providers drop old models when traffic falls, and model deprecation and migration covers how to have the replacement tested before the notice arrives.