GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Ismantic
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Ismantic
Model Tree
Input Modalities
Output Modalities
Supported Functionality
S0 — monolingual only. 12B tokens, Chinese/English 50:50, 45,149 steps. No parallel or instruction data at any point.
This is a 13B-token model. For scale: Qwen3-0.6B-Base saw 36T tokens —
about 2,700x more. Treat the numbers below as what that budget buys, not as a
competitive result.
| WMT22 5-shot | BLEU | COMET |
|---|---|---|
| zh->en | 0.54 | 0.4638 |
| en->zh | 3.97 | 0.5872 |
Few-shot translation is essentially zero, and that is the finding. The model ignores the in-context examples entirely — on zh->en it does not even switch output language. Language modelling was learned; in-context learning was not. 12B monolingual tokens is not enough for ICL to emerge at 0.5B.
Its value is as (a) the control for S1, and (b) a starting point for mid-training / SFT.
The tokenizer is a compiled C++ extension, not loadable by
AutoTokenizer. The release ships tokenizer.py and example_load.py:
bash
pip install git+https://github.com/Ismantic/PieceTokenizerpython example_load.py
The model code (model.py, checkpoint.py) is bundled too — the package
depends only on torch plus the tokenizer extension, not on transformers.
Full pipeline, data mixes and every design decision (including the mistakes)
are documented in https://github.com/Ismantic/Summer. Notably
docs/WHY.md records why fp32 master weights are mandatory, why the learning
rate schedule is WSD rather than cosine, and what the vocabulary swap cost.
Apache-2.0. Training corpora are public datasets (FineWeb-Edu, Cosmopedia,
CCI3-HQ, SkyPile, WMT19, OPUS-100 and others; see data/source.py upstream).
Please observe their respective licenses.