Instructions to use droplychee/droplychee-2.0-40b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use droplychee/droplychee-2.0-40b with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="droplychee/droplychee-2.0-40b")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("droplychee/droplychee-2.0-40b")
model = AutoModelForMultimodalLM.from_pretrained("droplychee/droplychee-2.0-40b", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use droplychee/droplychee-2.0-40b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "droplychee/droplychee-2.0-40b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "droplychee/droplychee-2.0-40b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/droplychee/droplychee-2.0-40b

SGLang

How to use droplychee/droplychee-2.0-40b with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "droplychee/droplychee-2.0-40b" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "droplychee/droplychee-2.0-40b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "droplychee/droplychee-2.0-40b" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "droplychee/droplychee-2.0-40b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use droplychee/droplychee-2.0-40b with Docker Model Runner:
```
docker model run hf.co/droplychee/droplychee-2.0-40b
```

🧠 Droplychee-2.0-40B

A multilingual long-context language model developed by Droplychee, built upon the Qwen family and extended through full fine-tuning and architectural modifications.

Overview

Droplychee-2.0-40B is a decoder-only transformer model designed for multilingual reasoning, software engineering, long-context understanding, and instruction following. The model is derived from the Qwen family and further adapted through full fine-tuning using curated multilingual instruction datasets.

The project focuses on delivering strong performance in English and Bangla while maintaining support for over 40 languages.

Key Features

🌍 Supports 40+ languages
📚 Maximum context length of 1,000,000 tokens
🧵 YaRN-based context extension
💻 Optimized for software engineering and coding tasks
🧠 Advanced reasoning and instruction following
📄 Long-document analysis and summarization
🤖 Agent-oriented workflows and tool use

Model Details

Property	Value
Model Name	Droplychee-2.0-40B
Organization	Droplychee
Base Model	DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking
Architecture	Decoder-only Transformer
Training Method	Full Fine-Tuning
Context Length	1,000,000 tokens
Context Extension	YaRN
Languages	40+
License	Apache-2.0

Architecture

Droplychee-2.0-40B is derived from the Qwen architecture with additional modifications introduced during development.

Key architectural characteristics include:

Decoder-only transformer architecture
Rotary Position Embeddings (RoPE)
YaRN-based long-context extension
Optimized KV-cache handling
Flash Attention support (where available in the serving stack)
Full supervised instruction fine-tuning

Training

The model was fine-tuned using a curated multilingual instruction corpus emphasizing:

General reasoning
Coding and software engineering
Mathematics
Multilingual dialogue
Long-context comprehension
Agent-oriented tasks

Approximate training statistics:

Item	Value
High-quality instruction pairs	~500K
Training tokens	~100M–1B
Training Platform	Unsloth Studio
GPU Hardware	3× NVIDIA RTX PRO 6000 Blackwell Server Edition

Context Extension

Droplychee-2.0-40B supports a maximum context length of 1,000,000 tokens through YaRN (Yet another RoPE extensioN).

YaRN extends the effective context window while preserving compatibility with Rotary Position Embeddings. Long-context performance may vary depending on the inference engine, hardware resources, and serving configuration.

Supported Languages

The model is optimized for multilingual instruction following and has been evaluated primarily on English and Bangla. It also supports more than 40 languages, including Hindi, Arabic, Chinese, Japanese, Korean, French, German, Spanish, Portuguese, and Russian.

Intended Use

Recommended applications include:

Conversational AI
Coding assistants
Research assistance
Educational tools
Long-document summarization
Retrieval-Augmented Generation (RAG)
AI agents
Software engineering workflows

Limitations

Outputs may contain factual inaccuracies or hallucinations.
Performance varies across languages and domains.
Independent third-party evaluation has not yet been completed.
Users should verify outputs before relying on them for high-impact decisions.

Responsible Use

This model is intended for research and general-purpose AI applications. Users are responsible for ensuring compliance with applicable laws, regulations, and ethical guidelines.

Attribution

This project is derived from:

DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking

Please refer to the original model card and license for upstream attribution requirements.

License

This model is distributed under the Apache License 2.0.

Users must also comply with the license terms of the upstream base model where applicable.

Evaluation

Note: The following results are based on internal evaluation conducted by the Droplychee team. Independent third-party verification has not yet been completed.

Reasoning Benchmarks

Benchmark	Score	Evaluation
MMLU	90.8	Internal
MMLU-Pro	89.5	Internal
GPQA Diamond	87.0	Internal
MGSM	90.4	Internal
Humanity's Last Exam	+11 pp	Internal
MMMU	80.7	Internal

Coding Benchmarks

Benchmark	Score	Evaluation
HumanEval	92.0	Internal
LiveCodeBench	76.8	Internal
SWE-Bench Verified	80.9	Internal
TerminalBench 2.0	59.3	Internal
SpreadsheetBench	64.25	Internal
SpreadsheetBench + Python	92.77	Internal
APEX-SWE	38.5	Internal

Agent Benchmarks

Benchmark	Score	Evaluation
τ² Telecom	98.2	Internal
τ² Retail	88.9	Internal
TerminalBench Hard	44.0	Internal
n8n AI Benchmark	66.0	Internal

Research Benchmarks

Benchmark	Score	Evaluation
Elicit Research Accuracy	96.5	Internal
Report Writing	62.0	Internal
METR Agency Benchmark	~5 Hours	Internal

Artificial Analysis

Benchmark	Score	Evaluation
Intelligence Index	70.0	Internal
Omniscience	10.0	Internal

Disclaimer: These benchmark results are derived from the Droplychee team's internal evaluation pipeline. Results may vary depending on hardware, inference engine, prompt format, evaluation methodology, and software versions. Independent third-party verification is planned for future releases.

For the latest documentation, technical reports, and official releases, please visit the official Droplychee GitHub repository. https://github.com/DropLychee/droplychee-2.0-40b

droplychee-2.0-40b/ │ ├── README.md ├── LICENSE ├── MODEL_CARD.md ├── TECHNICAL_REPORT.md ├── SYSTEM_CARD.md ├── EVALUATION.md ├── CITATION.cff ├── CHANGELOG.md ├── CONTRIBUTING.md ├── SECURITY.md ├── CODE_OF_CONDUCT.md ├── docs/ │ ├── architecture.md │ ├── training.md │ ├── datasets.md │ ├── tokenizer.md │ ├── inference.md │ ├── benchmarks.md │ ├── deployment.md │ ├── safety.md │ ├── governance.md │ └── roadmap.md ├── assets/ │ ├── logo.png │ ├── banner.png │ ├── architecture.svg │ └── benchmark_charts/ ├── examples/ │ ├── transformers.py │ ├── vllm.py │ ├── llama_cpp.py │ └── openai_api.py ├── scripts/ │ ├── evaluate.py │ ├── benchmark.py │ ├── convert_gguf.py │ └── export.py └── paper/ ├── Droplychee-2.0-40B_Technical_Report.pdf ├── Droplychee-2.0-40B_Technical_Report.md └── references.bib