Points clés
- Best overall coding model: Qwen2.5-Coder 32B — 87% HumanEval, requires 20 GB RAM at Q4_K_M.
- Best for 8 GB RAM: Qwen2.5-Coder 7B — 72% HumanEval, runs at 15–25 tok/sec on CPU.
- Best for fill-in-the-middle (code completion): Starcoder2 15B — purpose-built for IDE-style autocomplete.
- Code-specific models score 5–15 percentage points higher on HumanEval than general-purpose models at the same parameter count.
- For AI coding assistant workflows (VS Code, Cursor), see Local LLMs for Coding Workflows.
What Makes a Local LLM Good for Coding?
Coding performance in local LLMs is measured primarily by HumanEval — a benchmark of 164 Python programming problems where the model must generate a correct function body. HumanEval pass@1 scores (percentage of problems solved on the first attempt) are the standard comparison metric.
Code-specific models are fine-tuned on large code corpora (GitHub, Stack Overflow, documentation) and often include fill-in-the-middle (FIM) training — the ability to complete code given both the preceding and following context, which is required for IDE autocomplete.
General-purpose models like Llama 3.1 8B score 72% on HumanEval, which is competitive. But dedicated coding models at the same size score 5–15% higher because their training data and fine-tuning prioritize code generation accuracy over general language tasks.
#1 Qwen2.5-Coder 32B — Best Overall Local Coding LLM
Qwen2.5-Coder 32B is the highest-performing locally-runnable coding model in 2026. It scores 87% on HumanEval and 79% on MBPP (another Python coding benchmark). It supports 40+ programming languages including Python, JavaScript, TypeScript, Java, C++, SQL, Rust, and Go.
At Q4_K_M quantization, it requires ~20 GB RAM — manageable on workstations and MacBooks with 24+ GB of unified memory. Response quality on complex multi-file refactoring and algorithm design tasks is competitive with GPT-4o Mini.
| Spec | Value |
|---|---|
| HumanEval score | 87% |
| MBPP score | 79% |
| RAM required (Q4_K_M) | ~20 GB |
| Context window | 128K tokens |
| Languages | 40+ programming languages |
| Ollama command | ollama run qwen2.5-coder:32b |
#2 DeepSeek-Coder V2 Lite 16B — Best for 16 GB RAM
DeepSeek-Coder V2 Lite is a 16B mixture-of-experts coding model from DeepSeek. Despite 16B active parameters, it achieves 81% HumanEval through its MoE architecture and requires ~10 GB RAM at Q4_K_M. It is the best coding model for machines with 16 GB RAM.
It supports fill-in-the-middle completion and handles multi-language codebases well. The Lite variant uses 2.4B active parameters per forward pass, making inference faster than a comparable dense 16B model.
| Spec | Value |
|---|---|
| HumanEval score | 81% |
| RAM required (Q4_K_M) | ~10 GB |
| Context window | 128K tokens |
| Architecture | Mixture of Experts (MoE) |
| Ollama command | ollama run deepseek-coder-v2:16b |
#3 Qwen2.5-Coder 7B — Best Coding Model for 8 GB RAM
Qwen2.5-Coder 7B scores 72% on HumanEval — matching the general-purpose Llama 3.1 8B while using ~4.7 GB RAM. For users with 8 GB RAM who want the best coding performance without sacrificing headroom for other applications, this is the recommended choice.
It includes FIM support for code completion tasks and is compatible with the Continue.dev VS Code extension for local AI coding assistance.
| Spec | Value |
|---|---|
| HumanEval score | 72% |
| RAM required (Q4_K_M) | ~4.7 GB |
| Context window | 128K tokens |
| FIM support | Yes |
| Ollama command | ollama run qwen2.5-coder:7b |
#4 Starcoder2 15B — Best for IDE Autocomplete
Starcoder2 15B from Hugging Face BigCode is purpose-built for fill-in-the-middle code completion — the pattern used by IDE autocomplete tools. It scores 67% on HumanEval but excels specifically on FIM tasks where context comes from both before and after the cursor position.
Starcoder2 is the recommended model when integrating a local LLM into a VS Code or JetBrains IDE via Continue.dev or Tabby. For chat-style code generation, Qwen2.5-Coder performs better.
| Spec | Value |
|---|---|
| HumanEval score | 67% |
| RAM required (Q4_K_M) | ~9 GB |
| FIM support | Yes (primary use case) |
| Training data | 619 programming languages |
| Ollama command | ollama run starcoder2:15b |
#5 Llama 3.1 8B — Best General-Purpose Fallback for Coding
If you already have Llama 3.1 8B installed and do not want to download a separate coding model, it scores 72% on HumanEval — identical to Qwen2.5-Coder 7B. For everyday coding tasks (writing functions, explaining code, debugging), the quality difference between Llama 3.1 8B and a dedicated coding model is marginal. Switch to a coding-specific model for complex algorithm tasks or large codebase refactoring.
HumanEval Benchmark: Best Local Coding LLMs Compared
| Model | HumanEval | MBPP | RAM | FIM |
|---|---|---|---|---|
| Qwen2.5-Coder 32B | 87% | 79% | 20 GB | Yes |
| DeepSeek-Coder V2 Lite 16B | 81% | 71% | 10 GB | Yes |
| Qwen2.5-Coder 7B | 72% | 68% | 4.7 GB | Yes |
| Starcoder2 15B | 67% | 54% | 9 GB | Yes (primary) |
| Llama 3.1 8B | 72% | 68% | 5.5 GB | No |
Which Local Coding LLM Should You Use?
- 8 GB RAM, coding focus: `ollama run qwen2.5-coder:7b` — best HumanEval per GB of RAM.
- 16 GB RAM: `ollama run deepseek-coder-v2:16b` — 81% HumanEval at only 10 GB RAM.
- 20+ GB RAM (best quality): `ollama run qwen2.5-coder:32b` — highest available HumanEval locally.
- IDE autocomplete in VS Code: Starcoder2 15B via Continue.dev — FIM-optimized for cursor position completion.
- Already running Llama 3.1 8B: skip downloading a separate model — coding quality is equivalent to Qwen2.5-Coder 7B for everyday tasks.
Sources
- DeepSeek Coder Model — Official documentation and HumanEval benchmarks
- Qwen2.5 Coder — Model card with coding performance data
- Starcoder2 15B — Fill-in-the-middle specialized model for code completion
Common Mistakes When Choosing Coding Models
- Using a general-purpose model when a coding-specialized model exists for 2–5× better code generation.
- Not testing on the programming language you actually use — coding model rankings vary by language.
- Expecting perfect code generation from 7B models — they require more prompt engineering than GPT-4o.