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

Libraries

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

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

pipe = pipeline("text-generation", model="Convergent-ch/Zynthos-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/Zynthos-1")
model = AutoModelForCausalLM.from_pretrained("Convergent-ch/Zynthos-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/Zynthos-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/Zynthos-1
# Run inference directly in the terminal:
llama cli -hf Convergent-ch/Zynthos-1

Install from WinGet (Windows)

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

Use Docker

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

LM Studio
Jan

vLLM

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

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Convergent-ch/Zynthos-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/Zynthos-1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

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

SGLang

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

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

Lemonade

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

Pull the model

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

Run and chat with the model

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

List all available models

lemonade list

Atomic Chat

Zynthos-1

Zynthos-1 is a 246M-parameter English chat model made by Convergent. It is the largest model in the Convergent family, above Pico 1, Tavik-1 and Cobalt-1, and was trained from scratch: own tokenizer, own training script, random initialization.

It is a small model. It writes fluent, well-structured English, but it often gets facts wrong and it cannot do arithmetic or write working code. Please read the limitations below before using it.

Quick start

Ollama

ollama create zynthos-1 -f Modelfile
ollama run zynthos-1

The Modelfile in this repository contains the chat template and stop tokens. Without it, Ollama guesses a wrong template and the output is garbage.

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("Convergent-ch/Zynthos-1")
model = AutoModelForCausalLM.from_pretrained("Convergent-ch/Zynthos-1")
msgs = [{"role": "user", "content": "Explain photosynthesis in simple words."}]
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=200, do_sample=True, temperature=0.5, top_p=0.9)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

Chat format: <|bos|><|user|>\n...<|end|>\n<|assistant|>\n...<|end|>\n. Generation stops at <|end|> (id 7) or <|eos|> (id 2).

Model


Architecture	Llama-style decoder (RMSNorm, SwiGLU, RoPE, grouped-query attention)
Parameters	245,928,960 (tied input/output embeddings)
Layers / hidden / MLP	18 / 1024 / 2816
Attention heads / KV heads	16 / 8
Context length	2048 tokens
Tokenizer	Byte-level BPE, 32,768 tokens, trained on FineWeb-Edu and chat text
Files	`model.safetensors` (fp32), `zynthos-1.gguf` (Q8_0, 263 MB)

Training

Trained on one NVIDIA H100 on Modal in about 5.4 hours.

Pretraining: 3.4B tokens of FineWeb-Edu (sample-10BT), 6,484 steps of 524,288 tokens. AdamW, peak learning rate 8e-4, 300 warmup steps, cosine decay, bf16 autocast.
Chat tuning: 0.3B tokens, 572 steps on SmolTalk (general subsets, no code or math subsets), mixed with about 3 % template chats about the model's own identity. Loss only on the assistant's replies. Peak learning rate 2e-4.

Validation loss on held-out FineWeb-Edu text:

Step	Validation loss
Pretraining 500	4.103
Pretraining 1,000	3.560
Pretraining 2,000	3.250
Pretraining 4,000	3.036
Pretraining 6,484 (end)	2.935
After chat tuning	3.207

The rise after chat tuning is expected. The model shifts from plain web text toward chat replies, so it predicts web text slightly worse. The chat training loss ended at 1.45.

Data volume is below the commonly cited compute-optimal ratio of ~20 tokens per parameter (3.7B tokens total versus ~4.9B). The budget was limited by the free monthly compute credits.

Benchmarks

Convergent 3.0, measured with Convergent Bench in Ollama. Every score is the best of 10 runs per benchmark (run 1 greedy, runs 2 to 10 sampled at temperature 0.7 with different seeds).

Benchmark	Zynthos-1	Cobalt-1
Agentic coding (Basis / Bonus)	0.0% / 0.0%	0.0% / 0.0%
Knowledge work (Basis / Bonus)	25.4% / 21.9%	10.3% / 5.8%
Question answering (Basis / Bonus)	32.0% / 34.0%	18.0% / 14.0%
Humanity's Current Exam	1.84%	2.98%
HTML coding	1.19%	0.0%

What it does well

Tested in Ollama (Q8_0, temperature 0.5) with 16 fixed questions:

Knows its identity: "My name is Zynthos. I was made by Convergent." It also answers its size correctly (about 250M parameters).
Writes fluent, structured English: lists, explanations, short poems.
Simple, well-known facts are often right, for example Paris as the capital of France. Explanations of common topics such as photosynthesis are mostly correct.
Fast: about 540 tokens per second on an RTX 3060 Ti.

Compared with Cobalt-1 (100M) on the same questions, the answers are longer, better structured and more often on topic.

Limitations

These failures come from the same 16-question test:

Facts: often wrong. It called Mercury the largest planet. It explained the blue sky with an invented etymology. It made up stories when asked who wrote Romeo and Juliet.
Math: fails. "What is 12 + 7?" and "What is 23 * 17?" produced long invented derivations with wrong results.
Code: cannot write working code, because code was excluded from training. A request for a Python function produced Java/C-like fragments that repeat.
Translation and other languages: English only. A German translation request failed.
Family: it does not reliably know what Pico 1, Tavik-1 and Cobalt-1 are, and described them as genes.
Length: answers are often longer than needed and can repeat themselves or count list items wrong.

Do not use Zynthos-1 for anything where correctness matters: medical, legal, financial, safety or factual reference. It has no safety tuning beyond its training data.

License

Apache 2.0. Training data: FineWeb-Edu (ODC-By 1.0) and SmolTalk (Apache 2.0).

Downloads last month: 51

Safetensors

Model size

0.2B params

Tensor type

F32

Convergent-ch
/

Zynthos-1