Qwen first-gen: strong benchmarks, stronger successor

What it does Qwen is Alibaba Cloud’s first-generation family of base and chat-tuned large language models, spanning 1.8B to 72B parameters. The chat variants handle summarization, translation, coding, math, and agent-style tool use, while the base models were pretrained on up to 3 trillion tokens of multilingual data with a stated emphasis on Chinese and English. The repository also hosts inference and fine-tuning recipes, including support for Int4 and Int8 quantization, LoRA, Q-LoRA, and deployment integrations for vLLM and FastChat.

The interesting bit The project claims its 72B model beats LLaMA 2-70B across all listed benchmarks and tops GPT-3.5 on seven of ten tasks—claims backed by published tables but, as always, worth verifying on your own workload. More concretely, most of the family supports 32K context windows (the 14B model is the outlier at 8K), and the whole line runs on Ascend 910 and Hygon DCU silicon in addition to the usual NVIDIA stack.

Key highlights

• Model sizes from 1.8B to 72B, each with base and chat checkpoints, plus Int4 and Int8 quantized variants
• Chat models support tool use, agent behavior, and code-interpreter modes
• 32K context length on most variants; the 14B model is limited to 8K
• Fine-tuning support via full-parameter tuning, LoRA, and Q-LoRA with published GPU memory requirements
• Includes demo integrations for WebUI, CLI, and an OpenAI-compatible API server

Caveats

• The repository is explicitly no longer actively maintained; Alibaba directs users to the Qwen2 repo for ongoing work
• RLHF-aligned checkpoints are mentioned but noted as “not released yet” in the README
• Benchmark comparisons against LLaMA 2 and GPT-3.5 are presented by the project authors, not independent evaluators

Verdict Worth a look if you need a frozen, well-documented reference implementation of a major Chinese-English LLM family, especially for academic reproducibility or hardware-specific deployment on Ascend/DCU. Skip it if you want an actively maintained codebase—follow Qwen2 instead.