SimpleLLM/kode-14b-GGUF

Kode — EU-Trained Coding Models

Kode is a family of instruction-tuned coding models built for real-world software engineering tasks. Fine-tuned on Qwen2.5-Coder using DPO + SFT with Claude-generated training samples on A100 GPUs.

Kode is the backbone of Kode CLI, an open-source local alternative to Claude Code.

Model	Parameters	VRAM	Best For
kode-14b	14B	~10 GB (Q8) / ~9 GB (Q4)	Consumer GPUs, fast iteration
kode-32b	32B	~19 GB (Q4)	Maximum quality, production use

Key Features

• 🇪🇺 Trained in the EU — DSGVO/GDPR compliant, no data leaves Europe
• 🔧 Tool-calling native — Trained specifically for file operations, shell commands, code search
• 🎯 Production code focus — Training data from real codebases, not synthetic benchmarks
• 📐 7 languages — Rust, Go, TypeScript, Python, C#, SQL, CSS/Tailwind
• 🏠 Runs locally — 14B fits on a single consumer GPU (RTX 3080+)

Supported Languages & Tasks

Languages

Rust • Go • TypeScript • Python • C# • PostgreSQL • CSS/Tailwind

Tasks

• Code generation — Complete functions, modules, and files from natural language
• Code refactoring — Improve existing code structure and performance
• Code review — Identify bugs, security issues, and improvements
• Tool calling — File I/O, shell commands, grep/search (Kode CLI integration)
• Code completion — Context-aware completions

Training Details

Base Model

Qwen2.5-Coder (14B and 32B variants)

Training Pipeline

1. SFT (Supervised Fine-Tuning) — Claude-generated training samples across 7 languages (~841 curated queries covering data structures, async, error handling, APIs, testing, and more)
2. DPO (Direct Preference Optimization) — Preference pairs from Claude evaluations of model outputs
3. Tool-call SFT — Specialized training for tool-calling patterns (read_file, write_file, bash_execute, grep, etc.)

Infrastructure

• GPU: NVIDIA A100 80GB (2× for 32B full fine-tune, 1× for QLoRA)
• Framework: Transformers + PEFT + TRL + Unsloth
• LoRA config (32B): r=64, alpha=128, dropout=0.05, targeting all attention + MLP projections
• Precision: bfloat16
• Sequence length: 4096 tokens

Training Data

• ~841 curated training queries across 7 programming languages
• Claude-generated reference solutions (chosen) vs. local model outputs (rejected) for DPO
• Bilingual prompts (English + German)

Usage

Ollama (Recommended)

# Install and run
ollama pull simplellm/kode-14b
ollama run simplellm/kode-14b

# Or the larger model
ollama pull simplellm/kode-32b
ollama run simplellm/kode-32b

Ollama API

curl http://localhost:11434/api/chat -d '{
  "model": "simplellm/kode-14b",
  "messages": [
    {"role": "user", "content": "Write a Rust function to find prime numbers using the Sieve of Eratosthenes"}
  ]
}'

🤗 Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "simplellm/kode-14b"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {"role": "system", "content": "You are a coding assistant. Respond with clean, production-ready code."},
    {"role": "user", "content": "Write a thread-safe LRU cache in Rust using Arc and Mutex"},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))

llama.cpp

# Download GGUF
wget https://huggingface.co/simplellm/kode-14b-GGUF/resolve/main/kode-14b-Q8_0.gguf

# Run
./llama-cli -m kode-14b-Q8_0.gguf -p "Write a Go HTTP server with middleware" -n 1024

Hosted Inference

Try Kode without downloading at SimpleLLM.eu — EU-hosted, GDPR-compliant inference API.

Quantized Versions

Variant	Size	Quality	Speed
kode-14b (FP16)	~28 GB	Baseline	Baseline
kode-14b-Q8	~15 GB	Near-lossless	~1.2× faster
kode-14b (Q4)	~9 GB	Good	~1.5× faster
kode-32b (native/FP16)	~64 GB	Best	Slowest
kode-32b-Q4	~19 GB	Very good	Fast

Benchmarks

🚧 Coming soon — We are running HumanEval, MBPP, MultiPL-E, and tool-calling benchmarks. Results will be published here.

Benchmark	kode-14b	kode-32b	Qwen2.5-Coder-14B (base)
HumanEval	TBD	TBD	TBD
MBPP	TBD	TBD	TBD
MultiPL-E (Rust)	TBD	TBD	TBD
Tool-call accuracy	TBD	TBD	N/A

Limitations

• Optimized for the 7 supported languages; may underperform on others
• 4096 token context window (inherited from training config)
• Tool-calling format is specific to Kode CLI's tool schema
• Training data is bilingual (EN/DE) — other languages may have reduced quality

License

Apache 2.0 (inherited from Qwen2.5-Coder)

Citation

@misc{kode2025,
  title={Kode: EU-Trained Coding Models for Real-World Software Engineering},
  author={Kevin and SimpleLLM Team},
  year={2025},
  url={https://huggingface.co/simplellm/kode-14b}
}

Links

• 🌐 SimpleLLM.eu — Hosted inference
• 💻 Kode CLI — Local coding assistant
• 🤗 All models — HuggingFace collection