by Alibaba (Qwen Team)

Qwen 2.5 Coder 32B.

code open weights workstation 33B params 128K ctx Transformer Quality 92.7
Cheapest input
$0.66/M
on Alibaba DashScope
Cheapest output
$1.0/M
on Alibaba DashScope
Fastest
30 tok/s
on OpenRouter
Smallest GPU
1× Nvidia RTX A5000
Capability snapshot

What it's best at.

Coding 92.7
Coding (MBPP) 90.2

Scores normalised against benchmark ceilings (100 = perfect). Coloured by tier — coral 80+ frontier, lavender 65+ strong, sage 50+ solid, slate below.

Benchmarks

Published scores.

Benchmark Score Source
MBPP 90.2 official ↗
HumanEval 92.7 official ↗
Description

About Qwen 2.5 Coder 32B.

Qwen 2.5 Coder 32B is Alibaba's open-weight specialist coding model — fine-tuned from Qwen 2.5 32B on 5.5T tokens of code. Tops HumanEval and MBPP at the 32B tier; competitive with much larger general-purpose models on coding-specific benchmarks. Released under Apache-2.0 (the only Qwen 2.5 variant under fully permissive licensing). Runs on a single H100 (FP16) or A6000 (INT4). Popular in self-hosted coding assistants (Continue.dev, Aider, Cline) as a Claude Sonnet alternative for cost-sensitive workflows.

Architecture

How it's built.

Architecture
Transformer
Trained on
5.5T tokens
169 tokens per parameter — near the Chinchilla optimum.
Knowledge cutoff
Aug 2024
103 days from cutoff to release.
Context window

How much it can remember.

128K tokens ≈ 96,000 English words
4K 32K 128K 1M
Max output per call: 8K tokens
Capabilities

What it can do.

· Vision input
· Audio input
· Video input
Function calling
· Tool use
· JSON mode
Streaming
Fine-tuning