ContextQMD
Back to Jan

Jan-Code-4B

docs/src/pages/docs/desktop/jan-models/jan-code-4b.mdx

0.7.93.9 KB
Original Source

import { Callout } from 'nextra/components'

Jan-Code-4B

Jan-Code-4B is a lightweight, code-tuned 4B parameter model built for fast local inference. Fine-tuned on Jan-v3-4B-base-instruct, it is designed for practical coding tasks with an emphasis on handling well-scoped subtasks reliably while keeping latency and compute requirements low.

<a href="jan://models/huggingface/janhq/Jan-code-4b-gguf" style={{ display: 'inline-flex', alignItems: 'center', gap: '8px', padding: '10px 20px', backgroundColor: '#000', color: '#fff', borderRadius: '8px', fontWeight: '600', fontSize: '15px', textDecoration: 'none', marginTop: '8px', }}

Open in Jan </a>

Overview

PropertyValue
Parameters4B
Base ModelJan-v3-4B-base-instruct (Qwen3-4B-Instruct-2507)
Fine-tuning focusCode generation, editing, refactoring, debugging
LicenseApache 2.0

Capabilities

  • Coding assistant: Code generation, editing, refactoring, and debugging
  • Agent workflows: Use as a fast worker/sub-agent in agentic setups (e.g., generating patches or tests)
  • Claude Code integration: Can replace the Haiku model in a Claude Code setup for a fully local coding workflow
<Callout type="tip"> Jan-Code-4B is designed to work as a drop-in local alternative to cloud coding models in agentic pipelines, keeping your code private and inference fast. </Callout>

Performance

Jan-Code-4B leads Jan-v3-base-instruct and the Qwen3-4B-Instruct-2507 base model across all three coding and reasoning benchmarks:

  • Aider (challenging Exercism tasks): 19.0 vs 18.0 vs 12.9
  • Livecode Bench v6 (real-world coding): 51.0 vs 45.8 vs 35.1
  • AIME25 (advanced math reasoning): 53.0 vs 47.0 vs 47.0

Requirements

  • Memory:
    • Minimum: 8GB RAM (with Q4 quantization)
    • Recommended: 16GB RAM (with Q8 quantization)
  • Hardware: CPU or GPU
  • API Support: OpenAI-compatible at localhost:1337

Using Jan-Code-4B

Quick Start

  1. Download Jan Desktop
  2. Select Jan-Code-4B from the model list
  3. Start coding — no additional configuration needed

Deployment Options

Using vLLM:

bash
vllm serve janhq/Jan-code-4b \
    --host 0.0.0.0 \
    --port 1234 \
    --enable-auto-tool-choice \
    --tool-call-parser hermes

Using llama.cpp:

bash
llama-server --model Jan-code-4b-Q8_0.gguf \
    --host 0.0.0.0 \
    --port 1234 \
    --jinja \
    --no-context-shift
yaml
temperature: 0.7
top_p: 0.8
top_k: 20

What Jan-Code-4B Does Well

  • Code generation: Write new functions, classes, and modules from natural language descriptions
  • Editing & refactoring: Modify existing code with targeted, reliable edits
  • Debugging: Identify and fix bugs in provided code snippets
  • Agentic subtasks: Fast, focused execution of scoped coding tasks within larger agent pipelines
  • Low latency: Runs efficiently on consumer hardware with fast response times

Limitations

  • Model size: 4B parameters limit handling of very large codebases or complex architectural decisions in a single context
  • General tasks: Optimized for coding — for general-purpose tasks, consider Jan-v3-4B-base-instruct

Available Formats

GGUF Quantizations

  • Q4_K_M: Good balance of size and quality
  • Q5_K_M: Better quality, slightly larger
  • Q8_0: Highest quality quantization (recommended)

Models Available

Community