GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Convergent-ch
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Convergent-ch
Model Tree
Input Modalities
Output Modalities
Supported Functionality
bash
ollama create zynthos-1 -f Modelfileollama run zynthos-1The Modelfile in this repository contains the chat template and stop tokens. Without it, Ollama guesses a wrong template and the output is garbage.
python
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).
| 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) |
Trained on one NVIDIA H100 on Modal in about 5.4 hours.
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.
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% |
Tested in Ollama (Q8_0, temperature 0.5) with 16 fixed questions:
Compared with Cobalt-1 (100M) on the same questions, the answers are longer, better structured and more often on topic.
These failures come from the same 16-question test:
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.
Apache 2.0. Training data: FineWeb-Edu (ODC-By 1.0) and SmolTalk (Apache 2.0).
ollama create zynthos-1 -f Modelfileollama run zynthos-1from 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))