holasoymalva/honey-code-7b

🍯 Honey-Code: High-Performance Reasoning Coding LLM & Agent

Honey-Code is a next-generation coding foundation model and autonomous coding agent. It is developed via Chain-of-Thought (CoT) Knowledge Distillation from compatible frontier reasoning LLM providers and refined using Reinforcement Learning with Rule-Based Sandbox Verification (GRPO).

Honey-Code delivers strong performance on agentic programming, multi-file refactoring, competitive programming, and automated debugging, matching and exceeding the performance of top-tier open-source coding models.

---

🚀 Key Highlights

• 🧠 Extended Chain-of-Thought: Native <think>...</think> internal reasoning engine for architecture design, algorithmic verification, and edge-case planning.
• 🛠️ Autonomous Agent Harness: Native multi-turn tool calling (view_file, replace_file_content, run_command, grep_search, lint_check).
• ⚡ 100% Sandbox-Verified Distillation: Every synthetic training pair has been verified against compilers (gcc, rustc, tsc) and unit test suites (pytest).
• 🦙 Ready for Ollama & GGUF: Instant local deployment on Mac (Apple Silicon M-series), Windows, and Linux.

---

📊 Benchmark Results

Benchmark	Honey-Code 7B	Honey-Code 14B	Qwen2.5-Coder-32B	DeepSeek-R1-Distill-Qwen-14B
HumanEval (Pass@1)	88.4%	92.6%	90.2%	89.8%
MBPP (Pass@1)	86.2%	90.1%	88.0%	87.4%
LiveCodeBench	44.8%	52.3%	46.5%	49.2%
SWE-bench Lite	38.2%	45.6%	33.4%	41.5%

---

💻 Quick Start with Ollama

# Run Honey-Code locally with Ollama
ollama run honey-code

Or using the built-in interactive terminal agent:

# Run the autonomous coding agent
honey-code --interactive

---

🐍 Usage with Transformers & vLLM

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "holasoymalva/honey-code-7b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

prompt = "<|im_start|>user\nWrite an async rate limiter in Python using Redis token bucket.<|im_end|>\n<|im_start|>assistant\n<think>\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=4096, temperature=0.6)
print(tokenizer.decode(outputs[0], skip_special_tokens=False))

---

📜 Citation & License

Apache-2.0 License. Developed with ❤️ by the Honey-Code Community.