Instructions to use Convergent-ch/Andromeda-1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Convergent-ch/Andromeda-1 with Transformers:

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

pipe = pipeline("text-generation", model="Convergent-ch/Andromeda-1")
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("Convergent-ch/Andromeda-1")
model = AutoModelForCausalLM.from_pretrained("Convergent-ch/Andromeda-1", 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

llama.cpp

How to use Convergent-ch/Andromeda-1 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 Convergent-ch/Andromeda-1
# Run inference directly in the terminal:
llama cli -hf Convergent-ch/Andromeda-1

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Convergent-ch/Andromeda-1
# Run inference directly in the terminal:
llama cli -hf Convergent-ch/Andromeda-1

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 Convergent-ch/Andromeda-1
# Run inference directly in the terminal:
./llama-cli -hf Convergent-ch/Andromeda-1

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 Convergent-ch/Andromeda-1
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Convergent-ch/Andromeda-1

Use Docker

docker model run hf.co/Convergent-ch/Andromeda-1

Llama new
LM Studio
Jan

vLLM

How to use Convergent-ch/Andromeda-1 with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/Convergent-ch/Andromeda-1

SGLang

How to use Convergent-ch/Andromeda-1 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 "Convergent-ch/Andromeda-1" \
    --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": "Convergent-ch/Andromeda-1",
		"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 "Convergent-ch/Andromeda-1" \
        --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": "Convergent-ch/Andromeda-1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use Convergent-ch/Andromeda-1 with Ollama:
```
ollama run hf.co/Convergent-ch/Andromeda-1
```
Unsloth Desktop
Docker Model Runner
How to use Convergent-ch/Andromeda-1 with Docker Model Runner:
```
docker model run hf.co/Convergent-ch/Andromeda-1
```

Lemonade

How to use Convergent-ch/Andromeda-1 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Convergent-ch/Andromeda-1

Run and chat with the model

lemonade run user.Andromeda-1-{{QUANT_TAG}}

List all available models

lemonade list

Atomic Chat

Request access to Andromeda-1

Andromeda-1 is shared on request only. Requests are reviewed by hand by Convergent.

Andromeda-1 is licensed under the Andromeda License (see LICENSE.md). You may download and run it locally and use its outputs. You may not fine-tune or otherwise modify it, redistribute it, rename it, or offer it as a hosted service.

Andromeda-1

Andromeda-1 is a 246M-parameter English language model made by Convergent. It is the stronger, exclusive sibling of Zynthos-1: Convergent trained Zynthos-1 (its own model, trained from scratch) further on step-by-step math, reasoning, code, question answering and web-search tool use. Access is granted on request.

Quick start

Ollama

ollama create andromeda-1 -f Modelfile
ollama run andromeda-1

Use the Modelfile from this repository. It contains the chat template, the tool-calling format and the sampling settings (repeat_penalty 1.0, because step-by-step solutions repeat numbers on purpose).

Web search: Andromeda-1 is registered as a tool model in Ollama. When a web_search tool is offered (for example with Ollama's web search switch, which needs an Ollama account and internet access), the model answers factual questions with a tool call and then answers from the results.

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("Convergent-ch/Andromeda-1")
model = AutoModelForCausalLM.from_pretrained("Convergent-ch/Andromeda-1")
msgs = [{"role": "user", "content": "What is 347 * 26?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True)["input_ids"]
out = model.generate(ids, max_new_tokens=700, do_sample=False)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

Hardware

	Minimum	Recommended
RAM	2 GB (the model needs about 0.5 GB)	4 GB or more
Disk	about 300 MB	1 GB
GPU	not required	any GPU with 1 GB VRAM or more

Measured speed: about 540 tokens per second on an RTX 3060 Ti (Q8_0 GGUF).

Model


Base	Zynthos-1 by Convergent (trained from scratch on 3.7B tokens)
Architecture	Llama-style decoder, 18 layers, hidden size 1024, 16 heads / 8 KV heads, SwiGLU, RoPE
Parameters	245,928,960 (tied embeddings)
Context	2,048 tokens
Tokenizer	Zynthos-1 BPE (32,768 tokens) with all digit merges removed, so every digit is one token
Files	`model.safetensors` (fp32), `andromeda-1.gguf` (Q8_0, 263 MB)

Training

Fine-tuned from Zynthos-1 on one RTX 3060 Ti in about 5.5 hours: 124.7M tokens (78.7M of them trained on), learning rate 1e-4 with cosine decay, bf16. Loss only on the model's own replies.

Data	Examples
Synthetic step-by-step arithmetic	70,000
Synthetic reasoning (sequences, ordering, days of the week, letter counting, units, ages, decimals)	40,000
Code: Python, JavaScript, C++, C, HTML (CodeFeedback, Magicoder OSS-Instruct, Convergent tasks)	42,451
Short factual answers (NQ-Open)	87,000
Web-search tool use (SQuAD paragraphs as search results)	30,000 + 6,000 direct answers
General chat (SmolTalk)	50,000
Educational text (FineWeb-Edu)	25,000
Identity and persona (wise, calm, a little epic)	9,200

Validation loss on held-out data of the same mix: 1.781 before, 0.770 after fine-tuning.

Evaluation

ConvergentBench (Convergent's own benchmark, best of 10 runs)

Category	Andromeda-1	Zynthos-1
QA	34.0 %	32.0 %
QA (bonus)	30.0 %	34.0 %
Knowledge Work	21.8 %	25.4 %
Knowledge Work (bonus)	16.8 %	21.9 %
Agentic Coding	7.1 %	0.0 % (1 run)
Agentic Coding (bonus)	21.6 %	0.0 % (1 run)
Math	11.1 %	6.3 %
Math (bonus)	9.0 %	5.9 %
Humanity's Current Exam	1.9 %	1.8 %
HTML Coding	1.7 %	1.2 %

The results vary a lot between runs (for example QA between 8 % and 34 %). Averages over the 10 runs are lower: QA 26.4 %, Knowledge Work 9.7 %, Agentic Coding (bonus) 9.7 %, Math 6.6 %.

Convergent tests on new problems (none of them in the training data)

Test	Zynthos-1	Andromeda-1
Arithmetic, 420 problems (answer line must match exactly)	0 %	86.9 %
Reasoning, 100 problems	0	78
HTML snippets with valid structure	0 / 30	30 / 30
Python tasks passing their tests	0 / 50	17 / 50
Python tasks with function names never seen in training	0 / 20	0 / 20
Web search called with a query, 20 factual questions	0 / 20	20 / 20
Arithmetic answered directly although search was offered	10 / 10	10 / 10

Limitations

Facts: it often invents facts, and general knowledge and reading comprehension are weaker than in Zynthos-1 (lower Knowledge Work score). For example it named the wrong capital of France in a test.
Code: only short, simple programs; new kinds of tasks usually fail.
Emojis and personality show up less often than intended.
English only.
Small model: do not use it for medical, legal, financial or safety decisions.

License

Andromeda License 1.0 (see LICENSE.md): local use and use of outputs are allowed; fine-tuning, redistribution, renaming and hosting as a service are not.

Training data sources, with thanks to their authors: FineWeb-Edu and SmolTalk (Hugging Face), CodeFeedback-Filtered-Instruction (m-a-p), Magicoder OSS-Instruct (ise-uiuc), NQ-Open (Google Research), SQuAD (Stanford). Each is used under its own license.

Downloads last month: 2

Safetensors

Model size

0.2B params

Tensor type

F32