Instructions to use MenteEAI/MenteE-Code-3B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use MenteEAI/MenteE-Code-3B with Transformers:

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

pipe = pipeline("text-generation", model="MenteEAI/MenteE-Code-3B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("MenteEAI/MenteE-Code-3B")
model = AutoModelForCausalLM.from_pretrained("MenteEAI/MenteE-Code-3B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use MenteEAI/MenteE-Code-3B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "MenteEAI/MenteE-Code-3B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MenteEAI/MenteE-Code-3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/MenteEAI/MenteE-Code-3B

SGLang

How to use MenteEAI/MenteE-Code-3B 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 "MenteEAI/MenteE-Code-3B" \
    --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": "MenteEAI/MenteE-Code-3B",
		"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 "MenteEAI/MenteE-Code-3B" \
        --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": "MenteEAI/MenteE-Code-3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use MenteEAI/MenteE-Code-3B with Docker Model Runner:
```
docker model run hf.co/MenteEAI/MenteE-Code-3B
```

MenteE Code 3B (v1)

MenteE Code is a coding-agent assistant built by MenteE. This is v1: a LoRA fine-tune of Qwen/Qwen2.5-Coder-3B-Instruct (Apache-2.0), merged into a standalone model and trained to behave as MenteE Code.

Ask it who it is and it will tell you: "I am MenteE Code, a coding assistant built by MenteE." The chat template injects the MenteE identity by default, so every client (transformers, vLLM, llama.cpp, Ollama) presents it correctly.

Training (all numbers measured)

Item	Value
Base	Qwen2.5-Coder-3B-Instruct (3.09B params)
Data	53,567 schema-validated agent trajectories (deduplicated from 78,212)
Sources	SWE-agent, OpenHands, R2E-Gym, NVIDIA SWE-Zero subsets + MenteE collector + identity Q&A
Method	LoRA r=8, α=16, all-linear targets, bf16, gradient checkpointing
Schedule	1,250 steps · 1 epoch · effective batch 16 · lr 2e-4 · seq len 4,096
Context	32,768 native · 131,072 via YaRN (factor 4.0, set in config.json)
Hardware	1× RTX 5090 (32 GB) · 5 h 32 m · ~$0.13
Final	loss 1.62 → 0.317 · token accuracy 91.5%

Benchmarks — measured, not estimated

HumanEval pass@1, greedy decoding, 50 problems, both models run on identical hardware with an identical harness (generated code executed against the official unit tests):

Model	pass@1
Qwen2.5-Coder-3B-Instruct (base)	94.0%
MenteE Code 3B v1	50.0%

We publish this gap on purpose. Failure analysis shows 24 of 25 v1 failures are the same artifact — systematically doubled closing brackets — caused by v1's training recipe (loss applied to all trajectory roles, including noisy tool outputs). It is a data-recipe issue, not a capability ceiling. The v2 recipe (assistant-only loss, identity oversampling, LoRA r=16) targets exactly this.

This is a 50-problem subset (~±14 pt margin) — directional, not leaderboard-grade.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

t = AutoTokenizer.from_pretrained("MenteEAI/MenteE-Code-3B")
m = AutoModelForCausalLM.from_pretrained("MenteEAI/MenteE-Code-3B", torch_dtype="auto", device_map="auto")
msgs = [{"role": "user", "content": "Write a Python function to reverse a linked list."}]
prompt = t.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
enc = t(prompt, return_tensors="pt").to(m.device)
out = m.generate(**enc, max_new_tokens=512)
print(t.decode(out[0][enc["input_ids"].shape[1]:], skip_special_tokens=True))

GGUF builds (for Ollama / llama.cpp) are published at MenteEAI/MenteE-Code-3B-GGUF.

Roadmap

v2: assistant-only loss, identity oversampling, LoRA r=16 → recover base-level coding quality
Flywheel: real user conversations are logged (opt-in), converted to training data, and folded into future versions
Eval: SWE-bench Verified subset with a proper agent harness

License & attribution

Apache-2.0, same as the base model. Built on Qwen2.5-Coder-3B-Instruct by the Qwen team — with gratitude. MenteE is not affiliated with Alibaba or Qwen.

Downloads last month: 98

Safetensors

Model size

3B params

Tensor type

BF16

Model tree for MenteEAI/MenteE-Code-3B

Base model

Qwen/Qwen2.5-3B

Finetuned

Qwen/Qwen2.5-Coder-3B

Finetuned

Qwen/Qwen2.5-Coder-3B-Instruct

Adapter

(94)

this model

Adapters

2 models