Curiosity · AI Model
Qwen 2.5 Coder 32B
Qwen 2.5 Coder 32B Instruct is Alibaba Cloud's November 2024 code-specialized open-weights model — a 32B dense transformer that matched GPT-4o on HumanEval and MBPP at release. Apache 2.0 licensed and single-H100 deployable, it became the default self-host choice for many coding copilot teams.
Model specs
- Vendor
- Alibaba
- Family
- Qwen 2.5 Coder
- Released
- 2024-11
- Context window
- 131,072 tokens
- Modalities
- text, code
- Input price
- $0.15/M tok
- Output price
- $0.15/M tok
- Pricing as of
- 2026-04-20
Strengths
- Apache 2.0 — fully permissive commercial use
- GPT-4o class HumanEval at a single-H100 footprint
- Broad language coverage — 92 programming languages in pretraining
- Strong fill-in-the-middle and repo-level generation quality
Limitations
- Trails newer reasoning coders on multi-step planning benchmarks
- Coding-only focus — general chat quality below Qwen 2.5 72B Instruct
- LiveCodeBench score lower than some closed frontier coders
- Smaller community than Code Llama for adjacent tooling
Use cases
- Self-hosted coding copilots for IDE and terminal
- Automated code review and refactoring pipelines
- Repository-scale code generation with 128K context
- Fine-tuning base for language-specific coding assistants
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| HumanEval | ≈92% | 2024-11 |
| MBPP | ≈90% | 2024-11 |
| LiveCodeBench | ≈31% | 2024-11 |
Frequently asked questions
Is Qwen 2.5 Coder 32B the best open coder?
As of early 2026 it's tied with DeepSeek Coder V2 at the top of open-weights coding benchmarks. Qwen wins on single-GPU footprint and HumanEval; DeepSeek wins on MoE inference economics and broader languages.
What sizes does Qwen 2.5 Coder come in?
The family covers 0.5B, 1.5B, 3B, 7B, 14B, and 32B. The 32B Instruct variant is the flagship; smaller sizes are popular for local IDE completion.
Can Qwen 2.5 Coder run on a laptop?
The 7B and 14B variants run well on high-end laptops via Ollama or llama.cpp. 32B needs a desktop GPU (24GB+) or a workstation-class machine for responsive inference.
Sources
- Qwen — Qwen 2.5 Coder announcement — accessed 2026-04-20
- Hugging Face — Qwen/Qwen2.5-Coder-32B-Instruct — accessed 2026-04-20