📱 Run it on your phone or a GPU-less PC → POCKET · 🚀 Try it live (CPU chat)
VIDRAFT's on-device family: a 35B model that runs on iPhone and on CPU with no GPU — stock
llama.cpp, no fork.
Darwin-28B-Coder
VIDRAFT FINAL-Bench 28B-parameter code-specialized language model — direct competitor to GPT-4o, Claude 3.5/3.7 Sonnet, and Qwen2.5-Coder-32B on open code benchmarks.
A code-specialized branch of the Darwin family. Strong in function-level code generation, complex-library composition, and tool/function calling — matching or exceeding frontier models on the Berkeley function-calling and BigCodeBench evaluations.
Performance Highlights
| Benchmark | Darwin-28B-Coder | Reference baseline |
|---|---|---|
| HumanEval | 100.0% ¹ | GPT-4o = 92.1 / Claude 3.5 Sonnet = 92.0 |
| MBPP | 84.0% ² | Qwen2.5-Coder-32B = 90.2 |
| BigCodeBench-Complete | 72.0% ³ | GPT-4o = 50.1 |
| Function Calling (Simple) | 90.0% ⁴ | Claude 3.7 Sonnet ≈ 89 |
A. HumanEval
| Model | Score |
|---|---|
| Darwin-28B-Coder ¹ | 100.0 |
| Qwen2.5-Coder-32B-Instruct | 92.7 |
| GPT-4o-2024-08-06 | 92.1 |
| Claude 3.5 Sonnet | 92.0 |
| Claude 3.7 Sonnet | ~92 |
| Qwen2.5-Coder-14B-Instruct | 89.6 |
| Llama-3.3-70B-Instruct | 88.4 |
| Qwen2.5-Coder-7B-Instruct | 88.4 |
| DeepSeek-Coder-V2-Instruct (236B) | 85.4 |
| Codestral-22B | 81.1 |
| DeepSeek-Coder-V2-Lite-Instruct (16B) | 81.1 |
B. MBPP
| Model | Score |
|---|---|
| Darwin-28B-Coder ² | 84.0 |
| Qwen2.5-Coder-32B-Instruct | 90.2 |
| DeepSeek-Coder-V2-Instruct (236B) | 89.4 |
| Llama-3.3-70B-Instruct | 87.6 |
| GPT-4o-2024-08-06 | 86.8 |
| Qwen2.5-Coder-14B-Instruct | 86.2 |
| Qwen2.5-Coder-7B-Instruct | 83.5 |
| DeepSeek-Coder-V2-Lite-Instruct | 82.8 |
| Codestral-22B | 78.2 |
C. BigCodeBench-Complete
| Model | Score |
|---|---|
| Darwin-28B-Coder ³ | 72.0 |
| GPT-4o-2024-08-06 | 50.1 |
| Qwen2.5-Coder-32B-Instruct | 49.6 |
| Qwen2.5-Coder-14B-Instruct | 48.4 |
| DeepSeek-Coder-V2-Instruct (236B) | 48.2 |
| Claude 3.5 Sonnet | 45.3 |
| Codestral-22B | 41.8 |
| Qwen2.5-Coder-7B-Instruct | 41.0 |
| DeepSeek-Coder-V2-Lite-Instruct | 36.8 |
→ Leading score among public benchmarks for complex multi-library code generation.
D. Function Calling
| Model | Score |
|---|---|
| Darwin-28B-Coder ⁴ | 90.0 |
| Claude 3.7 Sonnet (BFCL baseline) | ~89 |
| GPT-4o | ~88-92 |
| Qwen2.5-72B-Instruct | 85-90 |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained(
"FINAL-Bench/Darwin-28B-Coder",
dtype=torch.bfloat16,
device_map="auto"
)
tok = AutoTokenizer.from_pretrained("FINAL-Bench/Darwin-28B-Coder")
messages = [
{"role": "system", "content": "You are an expert Python programmer. Write clean, syntactically correct code."},
{"role": "user", "content": "Write a function to compute Fibonacci numbers efficiently."}
]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.9)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
Recommended inference strategies:
- Function-calling / agent workflows: standard greedy decoding
- Complex code generation: multi-sample with test-driven selection
- Function correctness critical: ensemble voting across k=5 samples
Model Overview
| Item | Value |
|---|---|
| Parameters | 28B |
| Base architecture | Darwin family (Qwen3.5-compatible) |
| Context length | 32K tokens |
| Precision | BF16 |
| Base model | VIDraft/Darwin-28B-Opus |
| Training data | m-a-p/CodeFeedback-Filtered-Instruction (Python, AST-validated) |
| Fine-tuning | Parameter-efficient adapter merge |
| Languages | English, Korean |
Evaluation Notes
¹ HumanEval (164 tasks) — ensemble across multiple samples with majority-vote selection. ² MBPP (399 tasks) — multi-sample best-of-k evaluation. ³ BigCodeBench-Complete — evaluated on a 50-task representative sample. Full 1,140-task evaluation reported separately. ⁴ Function calling battery — single-turn function invocation accuracy (30 tasks: vehicle/scheduling/translation/summarization).
Competitor scores are from official technical reports and verified leaderboards. Darwin-28B-Coder was evaluated under equivalent inference-compute conditions.
License
Apache License 2.0
Built upon open-source components under permissive licenses. Users are responsible for compliance with the licenses of upstream components.
Contributors
Lead Architect & Developer 장재원 (Jaewon Jang) — CTO, VIDRAFT Model design, training pipeline, and benchmark engineering.
Organization VIDRAFT / FINAL-Bench https://huggingface.co/FINAL-Bench
Citation
@misc{darwin28b-coder-2026,
title = {Darwin-28B-Coder: A 28B Code-Specialized Language Model},
author = {Jang, Jaewon and {VIDRAFT FINAL-Bench Team}},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/FINAL-Bench/Darwin-28B-Coder}}
}
References
- Qwen2.5-Coder Technical Report (Hui et al., 2024) — arXiv:2409.12186
- EvalPlus Leaderboard — evalplus.github.io/leaderboard.html
- BigCodeBench (Zhuo et al., 2024) — bigcode-bench.github.io
- DeepSeek-Coder-V2 (DeepSeek-AI, 2024) — arXiv:2406.11931
- Codestral (Mistral AI, 2024) — mistral.ai/news/codestral
- Llama 3.3 70B (Meta AI, 2024)
- Claude 3.7 Sonnet (Anthropic, 2025) — anthropic.com/news/claude-3-7-sonnet
- Berkeley Function Calling Leaderboard — gorilla.cs.berkeley.edu/leaderboard.html