Instructions to use keyvan-ai/Mankei-1B-Chat with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use keyvan-ai/Mankei-1B-Chat with Transformers:

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

pipe = pipeline("text-generation", model="keyvan-ai/Mankei-1B-Chat")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("keyvan-ai/Mankei-1B-Chat")
model = AutoModelForCausalLM.from_pretrained("keyvan-ai/Mankei-1B-Chat", 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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use keyvan-ai/Mankei-1B-Chat with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf keyvan-ai/Mankei-1B-Chat:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf keyvan-ai/Mankei-1B-Chat:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf keyvan-ai/Mankei-1B-Chat:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf keyvan-ai/Mankei-1B-Chat:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf keyvan-ai/Mankei-1B-Chat:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf keyvan-ai/Mankei-1B-Chat:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf keyvan-ai/Mankei-1B-Chat:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf keyvan-ai/Mankei-1B-Chat:Q4_K_M

Use Docker

docker model run hf.co/keyvan-ai/Mankei-1B-Chat:Q4_K_M

LM Studio
Jan

vLLM

How to use keyvan-ai/Mankei-1B-Chat with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "keyvan-ai/Mankei-1B-Chat"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "keyvan-ai/Mankei-1B-Chat",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/keyvan-ai/Mankei-1B-Chat:Q4_K_M

SGLang

How to use keyvan-ai/Mankei-1B-Chat 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 "keyvan-ai/Mankei-1B-Chat" \
    --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": "keyvan-ai/Mankei-1B-Chat",
		"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 "keyvan-ai/Mankei-1B-Chat" \
        --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": "keyvan-ai/Mankei-1B-Chat",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use keyvan-ai/Mankei-1B-Chat with Ollama:
```
ollama run hf.co/keyvan-ai/Mankei-1B-Chat:Q4_K_M
```

Unsloth Studio

How to use keyvan-ai/Mankei-1B-Chat with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for keyvan-ai/Mankei-1B-Chat to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for keyvan-ai/Mankei-1B-Chat to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for keyvan-ai/Mankei-1B-Chat to start chatting

Docker Model Runner
How to use keyvan-ai/Mankei-1B-Chat with Docker Model Runner:
```
docker model run hf.co/keyvan-ai/Mankei-1B-Chat:Q4_K_M
```

Lemonade

How to use keyvan-ai/Mankei-1B-Chat with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull keyvan-ai/Mankei-1B-Chat:Q4_K_M

Run and chat with the model

lemonade run user.Mankei-1B-Chat-Q4_K_M

List all available models

lemonade list

Atomic Chat
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Mankei-1B-Chat (experimentell)

Deutsches Chat-Sprachmodell mit 1,01 Mrd. Parametern, von Grund auf in Deutschland trainiert — eigener deutscher Tokenizer (32k), auf einer einzelnen Karte. Gebaut für den CPU-Betrieb, und die Grundlage der Mankei-326M-Embedder- und Mankei-326M-Reranker-Modelle.

Fähigkeiten

Deutsche Faktenabfragen (z. B. „Hauptstadt von Bayern" → „München")
Rückfrage bzw. Zurückhaltung bei Unsicherheit statt erfundener Antworten
Flüssiges, grammatisches Deutsch

Benchmark — deutsches MMLU (deutsche From-Scratch-Modelle)

Multiple-Choice per Log-Likelihood auf deutschem MMLU, verglichen mit anderen von Grund auf trainierten deutschen Modellen (faire Klasse):

Modell	Größe	Accuracy
Mankei-1B-Chat	1,01 B	23,0 %
gpt2-xl-wechsel-german	1,56 B	20,2 %
LLäMmlein-1B	1,10 B	20,0 %
german-gpt2 (dbmdz)	0,12 B	20,0 %

Mankei-1B führt das deutsche From-Scratch-Feld an. Kleine From-Scratch-Modelle liegen bei reinem Faktenwissen generell nahe dem Zufallsniveau (25 %) — Mankeis Stärke sind deutscher Sprachfluss und souveräner lokaler Betrieb, nicht der Wissens-Benchmark.

Verwendung

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
tok = AutoTokenizer.from_pretrained("keyvan-ai/Mankei-1B-Chat")
model = AutoModelForCausalLM.from_pretrained("keyvan-ai/Mankei-1B-Chat", dtype=torch.bfloat16)
msgs = [{"role": "user", "content": "Was ist die Hauptstadt von Bayern?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, max_new_tokens=64, temperature=0.7, top_p=0.9,
                     repetition_penalty=1.15, do_sample=True)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

Empfohlenes Sampling: temperature 0.7 · top_p 0.9 · repetition_penalty 1.15.

Lokaler Betrieb (GGUF)

GGUF-Quantisierungen (f16 → Q3_K_M) direkt in diesem Repo — und gespiegelt im eigenen keyvan-ai/Mankei-1B-Chat-GGUF (Model-Tree):

Quantisierung	Größe	Download
f16	1,9 GB	Mankei-1B-Chat-f16.gguf
Q8_0	1,0 GB	Mankei-1B-Chat-q8_0.gguf
Q6_K	854 MB	Mankei-1B-Chat-q6_k.gguf
Q5_K_M	727 MB	Mankei-1B-Chat-q5_k_m.gguf
Q4_K_M	637 MB	Mankei-1B-Chat-q4_k_m.gguf
Q3_K_M	512 MB	Mankei-1B-Chat-q3_k_m.gguf

# Ollama
ollama run hf.co/keyvan-ai/Mankei-1B-Chat:Q4_K_M

# llama.cpp
llama-cli -hf keyvan-ai/Mankei-1B-Chat:Q4_K_M

Empfohlenes Sampling: temperature 0.7 · top_p 0.9 · repetition_penalty 1.15.

Architektur

Llama-Architektur · 1,013 Mrd. Parameter · hidden 1536 · 32 Layer · 24/6 Heads (GQA) · SwiGLU · RMSNorm · RoPE θ 500 000 · Kontext 4096 · Vokabular 32 768 · tied embeddings.

Trainingsdaten

Pretraining auf lizenzgeprüften deutschen und englischen Quellen (EN gezielt für Mathematik, Wissenschaft, Bildung):

Quelle	Domäne	Lizenz
FineWeb2-HQ	DE Web (Top-Filter)	ODC-By 1.0
FinePDFs	DE PDF-Volltext	ODC-By 1.0
FineMath	EN Mathematik	ODC-By 1.0
OpenWebMath	EN Mathematik	ODC-By 1.0
FineWeb-Edu	EN Web (Bildung)	ODC-By 1.0
Wikipedia (de)	DE Enzyklopädie	CC-BY-SA 4.0
German Commons (openlegaldata / openalex)	DE Recht & Wissenschaft	≥ CC-BY-SA 4.0

Chat-SFT: selbst erzeugt und CC0, gegen ein festes Prüfset dekontaminiert.

Downloads last month: -

Safetensors

Model size

1B params

Tensor type

BF16

Model tree for keyvan-ai/Mankei-1B-Chat

Quantizations

1 model

keyvan-ai
/

Mankei-1B-Chat