Qwen 2.5 Coder 1.5B
Lightweight Local Coding Model from Alibaba Cloud
Released September 2024 by Alibaba Cloud's Qwen team. Qwen 2.5 Coder 1.5B is a genuinely small code-focused LLM: the Instruct model scores 70.7% on HumanEval in Qwen's technical report with only 1.54 billion parameters. The Ollama download is 1.0GB (Q4) and runs in about 2GB, it was pretrained on source code in 92 programming languages, and it is licensed under Apache 2.0. One of the most accessible local coding models for resource-constrained hardware.
Model Specifications
Architecture
Context and Languages
Resource Requirements
Published Benchmark Results
HumanEval and MBPP Scores (from Qwen Technical Report)
Scores for Qwen2.5-Coder-1.5B-Instruct, the build Ollama ships, from the Qwen2.5-Coder Technical Report (Tables 16-17, read November 2, 2026). Vendor-reported, not independently re-run.
Code Generation Benchmarks
Context: What These Scores Mean
70.7% HumanEval means the Instruct model solves about 71 of the 164 HumanEval problems on its first attempt. The base model, before instruction tuning, scores 43.9%. For a 1.5B model, both are strong.
For comparison: in the same report, Qwen2.5-Coder-7B-Instruct scores 88.4% on HumanEval and DeepSeek-Coder-1.3B-Instruct 65.9%. HumanEval is short single-function Python, so the gap to 7B models grows on real multi-file work.
Practical impact: Useful for code completion, simple function generation, and boilerplate tasks. For complex multi-file reasoning, a larger model is still better.
HumanEval Scores: Small Base Models
Performance Metrics
Bar chart: base models, Qwen2.5-Coder Technical Report Table 5 (CodeGemma 2B from Google's CodeGemma model card). Radar: Qwen2.5-Coder-1.5B-Instruct, Tables 16-17.
VRAM and Quantization Options
VRAM Usage by Quantization Level
One of the main advantages of a 1.5B model is genuinely low resource usage. These are the Ollama download sizes for each quantization level; add the KV cache (about 28KB per token at FP16, so 0.1GB at a 4K context and 0.9GB at the full 32K). Most users should start with Q4_K_M, the default tag.
Memory Usage Over Time
Sizes from the Ollama registry, read November 2, 2026. Pull a specific build with, for example, ollama pull qwen2.5-coder:1.5b-instruct-q8_0.
92 Language Support
According to the Qwen2.5-Coder technical report, the pretraining source code spans 92 programming languages. Performance varies significantly by language: the 1.5B Instruct model averages 56.7% across the report's MultiPL-E languages. Languages with more training data (Python, JavaScript, TypeScript, Java, C++) produce better results; less common languages will have weaker output quality.
Strongest Languages
Most training data, best results
Decent Support
Good for common tasks
Limited Support
Less training data, weaker output
Note: The "92 languages" figure describes the pretraining data, not a quality guarantee. A 1.5B model is most useful in the widely used languages; for less common languages, expect simpler completions and more errors. Larger models like Qwen 2.5 Coder 7B handle the long tail of languages much better.
Hardware Compatibility for Qwen 2.5 Coder 1.5B
System requirements and compatible hardware configurations for running Qwen 2.5 Coder 1.5B locally
Budget Setup
- • Phi-3 Mini (3.8B)
- • Basic tasks
- • Good for learning
- • Cost: $500-800
Recommended
- • Llama 3.1 8B
- • Professional use
- • Great balance
- • Cost: $800-1500
Enthusiast
- • Llama 3.1 70B
- • Enterprise ready
- • Maximum quality
- • Cost: $1500-3000
Local Coding Model Alternatives
| Model | Ollama tag | Download | Speed ceiling (M3 Pro, 150 GB/s) | HumanEval |
|---|---|---|---|---|
| Qwen 2.5 Coder 1.5B | qwen2.5-coder:1.5b | 0.99 GB | ~152 tok/s ceiling | 43.9% (base) / 70.7% (Instruct) |
| CodeGemma 2B | codegemma:2b | 1.55 GB | ~97 tok/s ceiling | 31.1% (base) |
| StarCoder2 3B | starcoder2:3b | 1.71 GB | ~88 tok/s ceiling | 31.7% (base) |
| DeepSeek Coder 1.3B | deepseek-coder:1.3b | 0.78 GB | ~192 tok/s ceiling | 34.8% (base) / 65.9% (Instruct) |
| Phi-2 2.7B | phi:2.7b | 1.60 GB | ~94 tok/s ceiling | not in the Qwen report |
Ceiling = memory bandwidth (Apple's 150 GB/s for the M3 Pro) ÷ file size: an upper bound, not a measurement. For scale, our one measured run on that machine, qwen2.5-coder:3b-base (1.93GB), generated 55.0 tok/s against a 78 tok/s ceiling (Ollama 0.40.13, 2 November 2026). Small models usually land further below their ceiling because fixed per-token overhead matters more.
Choosing the Right Small Coding Model
When to Choose Qwen 2.5 Coder 1.5B
- - You want the best published HumanEval score in this size class
- - You want Apache 2.0 licensing with no restrictions
- - Your VRAM budget is 1-2GB
- - You primarily work with Python, JS, or TypeScript
- - You need a 32K context window
When to Consider Alternatives
- - StarCoder2 3B: Trained on The Stack v2; lower base HumanEval (31.7%) at twice the size
- - Phi-2 2.7B: Better general reasoning, also good at code
- - DeepSeek Coder 1.3B: Even smaller, but weaker benchmarks
- - CodeGemma 2B: Google's option, good for fill-in-middle
- - Any 7B model: If you have 4-8GB VRAM, much better quality
Installation and Setup
System Requirements
Install Ollama
Download and install the Ollama runtime for your platform
Pull Qwen 2.5 Coder 1.5B
Download the model (approximately 1.0GB for the default Q4 quantization)
Run the Model
Start an interactive coding session
Optional: Limit Resource Usage
Useful for memory-constrained systems or running alongside other tools
Practical Use Cases
Good Use Cases
- - Code completion: Inline suggestions in editors like VS Code with Continue
- - Simple function generation: Boilerplate, utility functions, unit tests
- - Resource-limited environments: Old laptops, Raspberry Pi, VMs with low memory
- - Offline development: No internet required after model download
- - Learning and experimentation: Low barrier to entry for trying local AI
Limitations to Be Aware Of
- - Complex reasoning: Multi-step logic problems are weak at 1.5B -- for math-heavy work, the specialised Qwen 2.5 Math 7B is a far better fit
- - Large codebase understanding: Cannot reason about full project architecture
- - Niche languages: Quality drops sharply outside top 20 languages
- - Debugging complex bugs: Often misidentifies root causes
- - Documentation generation: Tends to produce generic or incomplete docs
Honest Assessment
What Qwen 2.5 Coder 1.5B Actually Is
Qwen 2.5 Coder 1.5B is genuinely impressive for a 1.5B parameter model. Its 70.7% HumanEval score (Instruct, vendor-reported) is ahead of older 1-3B coding models in Qwen's own comparison tables.
However, it is important to be clear: it does not match current 7B coding models. Qwen2.5-Coder-7B-Instruct scores 88.4% on HumanEval in the same report, and the gap is larger on multi-file, real-repository work than on HumanEval's single functions. If you have the hardware to run a 7B model (a 4.7GB download for qwen2.5-coder:7b), you will get better results.
The real value of this model is accessibility:
- - It runs on hardware that cannot load a 7B model at all
- - It leaves enough VRAM headroom for other applications alongside it
- - It provides useful (not great) code assistance at minimal resource cost
- - The Apache 2.0 license makes it usable anywhere without legal concerns
Bottom line: Choose Qwen 2.5 Coder 1.5B when you need a coding model that fits where a 7B model cannot. Choose a 7B model when you can afford the resources.
Frequently Asked Questions
What are the real HumanEval and MBPP scores for Qwen 2.5 Coder 1.5B?
The Qwen2.5-Coder technical report (arXiv 2409.12186) lists the base model at 43.9% on HumanEval and 69.2% on MBPP, and the Instruct model, which is what ollama pull qwen2.5-coder:1.5b downloads, at 70.7% on HumanEval, 66.5% on HumanEval+ and 69.2% on MBPP. These are vendor-reported numbers. Earlier versions of this page quoted 43.3% and 50%; those figures were wrong.
How much VRAM does Qwen 2.5 Coder 1.5B need?
Ollama's downloads are 1.0GB at Q4_K_M (the default tag), 1.1GB at Q5_K_M, 1.6GB at Q8_0 and 3.1GB at FP16. Add the KV cache, about 28KB per token at FP16: roughly 0.1GB at a 4K context and 0.9GB at the full 32K. In practice the Q4 build runs in about 2GB. It also runs on CPU-only systems, more slowly.
Does Qwen 2.5 Coder 1.5B really support 92 programming languages?
According to the Qwen2.5-Coder technical report, the pretraining source code spans 92 programming languages. Performance varies significantly by language -- Python, JavaScript, and TypeScript get the strongest results, while less common languages will have weaker output. In practice, expect useful results mainly in widely used languages.
How does it compare to StarCoder2 3B and DeepSeek Coder 1.3B?
In the Qwen2.5-Coder technical report's base-model table, Qwen 2.5 Coder 1.5B scores 43.9% on HumanEval, StarCoder2 3B 31.7% and DeepSeek Coder 1.3B 34.8%. So the Qwen model beats StarCoder2 3B at half the size and DeepSeek Coder 1.3B at a similar size. Among Instruct models it scores 70.7% against 65.9% for DeepSeek Coder 1.3B Instruct.
What is the context window for Qwen 2.5 Coder 1.5B?
The model card lists a full context length of 32,768 tokens (32K). The YaRN extension to 128K that Qwen documents applies to the 7B and larger Qwen2.5-Coder models, not the 1.5B. For code completion and short generation tasks, 32K is plenty.
What license is Qwen 2.5 Coder 1.5B released under?
Qwen 2.5 Coder 1.5B is released under the Apache 2.0 license, which is fully permissive for both commercial and non-commercial use. You can use it freely in production applications, modify it, and distribute it without restriction. This makes it one of the most legally accessible small coding models available.
Get Started with Qwen 2.5 Coder 1.5B
If you have limited VRAM and need a local coding assistant, Qwen 2.5 Coder 1.5B is one of your best options. It is free, permissively licensed, and genuinely small enough to run on almost anything.
ollama run qwen2.5-coder:1.5bApache 2.0 license -- free for any use -- 1.0GB download at Q4 -- 32K context
Build Real AI on Your Machine
RAG, agents, NLP, vision, and MLOps - chapters across 25 courses that take you from reading about AI to building AI.
Go from reading about AI to building with AI
25 structured courses. Hands-on projects. Runs on your machine. Start free.
Written by
Barsha Mohapatra
Founder and Editor, Local AI Master · Founder, Little AI Master
Founder of Little AI Master and the editor behind every course and guide on Local AI Master. Years of tutoring students taught her where AI education actually fails.
Related Guides
Continue your local AI journey with these comprehensive guides
- PILLARLocal AI Hardware Requirements (2026): Complete Guide
- Best Mac for Local AI: Every Apple Silicon Chip Ranked M1–M6
- Ollama System Requirements: Minimum RAM, GPU, Disk Size
- Run an LLM on Your Phone: Offline AI on Android & iPhone
- DeepSeek V4 Hardware Requirements: Flash vs Pro
- Best GPU for AI Video Generation: By VRAM Tier (2026)
- Wan 2.2 VRAM Requirements by GPU: 8GB to 24GB Guide
- DGX Spark vs Strix Halo vs Mac Studio for LLMs
- Running LLMs on CPU Only: What Actually Works Without a GPU
- Dual GPU vs One Big GPU for Local LLMs: What Pools
Grab the AI Starter Kit — career roadmap, cheat sheet, setup guide
No spam. Unsubscribe with one click.
Found your model? Now build something with it.
25 hands-on courses — RAG, agents, fine-tuning — all running locally. First chapter free, no card.