Claude-Coder-3B: Native Local Model for Claude Code CLI

Claude-Coder-3B is a fine-tuned 3B parameter coding specialist specifically trained and quantized to run vanilla, unmodified Claude Code CLI locally on consumer GPUs with under 4 GB VRAM (such as NVIDIA GeForce GTX 1650, RTX 3050, or modern integrated GPUs) natively inside Ollama—with zero external proxy bridges or middleware.

---

🎯 Motivation

Claude Code is one of the most capable agentic coding CLIs available, but by default it routes to remote Anthropic APIs. Existing local setups typically require heavy 14B–32B models or complicated Python proxy middleware that crash on laptop GPUs.

Claude-Coder-3B solves this: - Runs directly inside Ollama’s native /v1/messages endpoint. - Fits 100% of all 35 transformer layers into 4 GB VRAM. - Emits clean tool calls that Ollama’s internal AST parser processes instantly. - Completely eradicates base model capability disclaimers (“I’m sorry, but I can’t create or run external website builds…”).

---

⚡ Quick Start Guide

Step 1: Pull the Model

Make sure Ollama is installed and running:

ollama run hisparshmishra1/claude-coder-3b

Step 2: Configure Claude Code Environment

Point Claude Code to your local Ollama instance:

Windows (PowerShell)

$env:ANTHROPIC_BASE_URL = "http://localhost:11434"
$env:ANTHROPIC_AUTH_TOKEN = "ollama"
$env:ANTHROPIC_MODEL = "hisparshmishra1/claude-coder-3b"

macOS / Linux (Bash or Zsh)

export ANTHROPIC_BASE_URL="http://localhost:11434"
export ANTHROPIC_AUTH_TOKEN="ollama"
export ANTHROPIC_MODEL="hisparshmishra1/claude-coder-3b"

Step 3: Launch in a Project Folder

Always cd into a dedicated project folder first (e.g., my-project):

mkdir my-app
cd my-app

claude --tools "Bash,Edit,Read,Write" --dangerously-skip-permissions

---

💻 How to Use This Model in Claude Code

Once your session launches, you can interact with Claude Code just like you would with remote models. Here are the most effective workflows:

1. Interactive Mode (Standard Workflow)

Launch Claude Code in your project directory:

claude --tools "Bash,Edit,Read,Write" --dangerously-skip-permissions

You will enter the interactive prompt (>). Now you can type instructions naturally:

• Building New Apps & Websites from Scratch:

  > Build a modern dark-mode calculator. Create index.html, styles.css, and app.js with full calculation logic.
  

  The model will automatically invoke Write to create each file on disk.

• Inspecting & Modifying Existing Code:

  > Inspect index.html and add a responsive navigation bar with links to Home, Features, and Contact.
  

  The model will use Read to inspect the current code, then use Edit to apply surgical modifications.

• Terminal Commands & Testing:

  > Run the test suite with npm test and fix any failing tests.
  

  The model will call Bash to run the tests, parse the failure output, and edit the code until tests pass.

• Conversations & Technical Questions:

  > How does the calculation logic in app.js work?
  

  The model will explain the code in clean GitHub-flavored markdown without triggering unnecessary tool calls.

---

2. Resuming Previous Conversations

If you close your terminal or step away, you can resume your exact conversation in that directory at any time:

claude --continue

(or simply claude -c)

---

3. Non-Interactive / Scripting Mode (-p)

To run a single automated command without entering the interactive shell:

claude -p "create a basic README.md explaining this project" --tools "Bash,Edit,Read,Write" --dangerously-skip-permissions

---

4. Useful In-Session Slash Commands

While inside an interactive Claude Code session, use these commands to optimize your workflow:

• /config: Configure session settings. Go here and disable Recaps to cut latency in half.
• /effort: Adjust thinking effort (low or normal is recommended for 3B models to maximize speed).
• /compact: Summarize and compress conversation history when your session grows long.
• /clear: Clear the screen and start a fresh context in the same directory.

---

💡 3 Essential Tips for Maximum Speed & Performance

On local consumer GPUs (like GTX 1650 / RTX 3050), apply these three optimizations for the best experience:

1. Disable Recaps (/config)

• What: Claude Code by default sends an extra background request after every single turn to generate a summary recap.
• How: In your Claude Code session, type /config, navigate to Recaps, and select Disable.
• Impact: Cuts your waiting time in half (from ~5–6 minutes down to ~1–2 minutes) and stops background token waste.

2. Limit Tools with --tools "Bash,Edit,Read,Write"

• What: Claude Code injects 10+ tool schemas (including web scrapers and notebook tools) by default, adding ~5,000 tokens of input context to every request.
• How: Always launch Claude Code with --tools "Bash,Edit,Read,Write".
• Impact: Drops input context by ~70%, speeding up prompt evaluation by 3x to 4x.

3. Never Launch in Root / Home Directories

• What: Claude Code scans the current working directory tree on startup.
• How: Always run inside a specific project folder (e.g., Desktop\my-project), never inside C:\Users\Username or ~.
• Impact: Prevents the context window from exploding to 18,000+ tokens and prevents HTTP connection timeouts (ECONNREFUSED).

---

📊 Hardware & Benchmark Specifications

Metric	Specification
Model Architecture	Qwen 2.5 Coder 3B Instruct
Fine-Tuning Method	LoRA (Rank 64, Alpha 128) across 4 balanced trajectory splits
Quantization Format	GGUF Q4_K_M via llama.cpp
Minimum VRAM	4 GB VRAM (GTX 1650, RTX 2050/3050/4050, Apple Silicon)
GPU Offload	100% GPU (35⁄35 Layers)
Context Window	32,768 tokens
W3C Standards Compliance	Valid HTML5 (Zero Quirks Mode warnings in Chrome DevTools)
License	Apache 2.0 / Qwen Research License