# Qwen2.5 Coder 32B Instruct ![Organization](https://img.shields.io/badge/Org-Alibaba-ff6a00) ![License](https://img.shields.io/badge/License-Apache%202.0-22c55e) ![Released](https://img.shields.io/badge/Released-Oct%202024-555) ![Focus](https://img.shields.io/badge/Focus-Coding-1f6feb) ![Params](https://img.shields.io/badge/Params-32B-blue) > **Quick answer:** Qwen2.5 Coder 32B Instruct is Alibaba's October 2024 code-specialised 32B model, scoring 92.7% HumanEval, 85.3% ACEBench, 91.1% GSM8K, and 30.8% BigCodeBench. Apache 2.0 — the largest Qwen2.5 Coder variant and strongest open-weight coder for its era. ## At a Glance **Where Qwen2.5 Coder 32B Instruct leads** - 92.7% HumanEval — excellent code generation - 85.3% ACEBench — strong agent/tool calling - 91.1% GSM8K — solid math - 83.0% HellaSwag — good language understanding - Apache 2.0 — fully open - 32B: more capacity than 7B Coder while staying efficient **Where it lags** - 30.8% BigCodeBench — complex code generation is limited - 27% BigCodeBench-Hard — difficult code problems are challenging - Released October 2024 — pre-dates Qwen3 series **Best for:** Coding assistants, code review, and code generation at 32B scale; open-source IDE integration; agent frameworks requiring strong tool use. ## What Qwen2.5 Coder 32B Instruct Is Qwen2.5 Coder 32B Instruct is the flagship model in the Qwen2.5 Coder series — the 32B parameter code-specialised variant released October 2024. It was the strongest open-weight coding model at its release, surpassing models like DeepSeek-Coder and CodeLlama on HumanEval. The 85.3% ACEBench score reflects solid agentic and tool-calling performance, making it suitable for coding agent frameworks. The 30.8% BigCodeBench indicates limitations on very complex code generation tasks — an area where larger reasoning models (QwQ-32B, DeepSeek-R1) have since improved. ## Specifications | Field | Value | |---|---| | Organization | Alibaba | | License | Apache 2.0 | | HuggingFace | Qwen/Qwen2.5-Coder-32B-Instruct | | Release date | October 7, 2024 | | Parameters | 32B | | Modality | Text / Code | | Context window | 128K tokens | ## Pricing Open weights under Apache 2.0 — self-host at no cost. Available via major inference providers. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [HumanEval](/benchmarks/humaneval) | 92.7% | Benchgen evaluation | 2024-10 | | [ACEBench](/benchmarks/acebench) | 85.3% | Benchgen evaluation | 2024-10 | | [GSM8K](/benchmarks/gsm8k) | 91.1% | Benchgen evaluation | 2024-10 | | [HellaSwag](/benchmarks/hellaswag) | 83.0% | Benchgen evaluation | 2024-10 | | [BigCodeBench](/benchmarks/bigcodebench) | 30.8% | Benchgen evaluation | 2024-10 | | [BigCodeBench Hard](/benchmarks/bigcodebench) | 27% | Benchgen evaluation | 2024-10 | ## Qwen2.5 Coder 32B vs Alternatives | Model | HumanEval | ACEBench | Params | License | |---|---|---|---|---| | **Qwen2.5 Coder 32B Instruct** | 92.7% | 85.3% | 32B | Apache 2.0 | | [Qwen2.5 Coder 7B Instruct](/models/alibaba/qwen2-5-coder-7b-instruct) | 88.4% | 49.6% | 7B | Apache 2.0 | | [Qwen2.5 72B Instruct](/models/alibaba/qwen2-5-72b-instruct) | 86.6% | — | 72B | Qwen | | [Kimi K2 Instruct 0905](/models/moonshot/kimi-k2-instruct-0905) | — | 76.5% | MoE | Apache 2.0 | Qwen2.5 Coder 32B vs 7B Coder: higher HumanEval (92.7% vs 88.4%) and much higher ACEBench (85.3% vs 49.6%) at ~4.6x the parameter count. Use 7B for lighter deployments, 32B for coding agents. ## Related Models & Benchmarks - Compare: [Qwen2.5 Coder 7B Instruct](/models/alibaba/qwen2-5-coder-7b-instruct) · [Kimi K2 Instruct 0905](/models/moonshot/kimi-k2-instruct-0905) · [Qwen2.5 72B Instruct](/models/alibaba/qwen2-5-72b-instruct) - Benchmarks: [HumanEval](/benchmarks/humaneval) · [ACEBench](/benchmarks/acebench) · [BigCodeBench](/benchmarks/bigcodebench) ## Frequently Asked Questions

What is Qwen2.5 Coder 32B Instruct?

Alibaba's October 2024 code-specialised 32B model scoring 92.7% HumanEval, 85.3% ACEBench, and 91.1% GSM8K. Apache 2.0 — the flagship Qwen2.5 Coder model.

Is Qwen2.5 Coder 32B better than Qwen2.5 Coder 7B?

Yes on most metrics: 92.7% vs 88.4% HumanEval, 85.3% vs 49.6% ACEBench, 91.1% vs 83.9% GSM8K. Use 7B for resource-constrained deployments; 32B for production coding agents.

--- *Specs from Alibaba's Qwen2.5 Coder 32B Instruct release (October 2024) and Benchgen evaluations. Last updated 2026-07-24.*

Qwen2.5-Coder 32B Instruct — Training jobs

Training jobs