Têtu-2
Têtu-2 is a small French language model trained from scratch with a custom architecture (October 2026). No pretrained weights were used. It comes with a per-word confidence estimate, shown as colors in the chat script, and a built-in calculator for the arithmetic it writes out.
Têtu-2 is a research model. It is small and often wrong, and it should not be used as a reliable assistant.
Versions
| Folder | Version | Description |
|---|---|---|
. (default) |
Têtu-2 chat (v2.1) | Conversational model: everyday French chat and step-by-step reasoning |
base/ |
Têtu-2 base | Pre-trained model (6.23B tokens), for text completion or further fine-tuning |
Model details
| Developer | tetu-AI |
| Architecture | Decoder-only transformer (custom) |
| Parameters | 702M total, 115M active per token (the rest is a lookup memory with negligible compute cost) |
| Context length | 32,768 tokens |
| Languages | French first, English |
| Vocabulary | 32,768 tokens (French and English, digits split individually) |
| Weights | bfloat16 (model.safetensors), doubt head in float32 |
| License | CC BY-NC 4.0 (non-commercial) |
Features:
- a doubt head that estimates, for every word, the probability that it is wrong (the chat script colors words green, yellow or red accordingly);
- a built-in calculator: when the model writes « 48 + 27 + 66 = », the exact result is inserted;
- about 275 tokens/s on an RTX 5070 (Têtu v1: 135).
Training
| Stage | Data |
|---|---|
| Pre-training | 6.23B tokens, 18.5 h on one A100: FineWeb-Edu, FineWeb2-HQ (French), French Wikipedia, FineMath, and program-verified synthetic data at the end |
| Instruction tuning (v2.0) | Program-verified generators (math, stories, multi-turn conversations that refer to earlier turns, concepts), French conversations and executed code |
| Chat tuning (v2.1) | Everyday French conversations written by Qwen2.5-7B-Instruct (Apache 2.0) as a teacher; reasoning answers kept only when two samples agreed or the code ran |
Evaluation
MGSM-fr
250 grade-school math problems translated into French by humans. All models were run on the same problems with greedy decoding and the same answer checker. These problems were not used in training.
| Model | Parameters | MGSM-fr |
|---|---|---|
| Qwen2.5-1.5B-Instruct | 1,544M | 48.4 % |
| Qwen2.5-0.5B-Instruct | 494M | 17.2 % |
| Têtu-2 chat (v2.1) | 115M active (702M total) | 5.6 % |
| Têtu v1.1 (previous model) | 96M active (181M total) | 4.8 % |
| Têtu-2 instruct (v2.0) | 115M active (702M total) | 4.0 % |
| SmolLM2-135M-Instruct | 135M | 1.6 % |
| MobileLLM-R1.5-140M | 140M | 0.4 % * |
* MobileLLM-R1.5-140M did not finish its answers within 768 tokens with our prompt, so this score underestimates it.
The Têtu-2 scores use the built-in calculator; the other models answer without one. Têtu v1 reaches 11.2 % with a 16-sample majority vote, which was not run for Têtu-2. Têtu-2 is ahead of models of similar size but well behind Qwen2.5-0.5B, which has about 4 times more active parameters and was trained on about 3,000 times more tokens.
Chat quality (pairwise judge)
Qwen2.5-14B-Instruct judged 250 held-out requests, seeing each pair of answers in both orders.
| Wins | Losses | Ties | p (sign test) | |
|---|---|---|---|---|
| Têtu-2 chat v2.1 vs Têtu v1.1 | 189 | 26 | 35 | 10⁻³¹ |
| Têtu-2 chat v2.1 vs Têtu-2 instruct v2.0 | 156 | 44 | 50 | 7·10⁻¹⁶ |
The judge belongs to the same model family as the chat teacher and may favor its style.
Synthetic test sets
Held-out phrasings from our own generators, with the calculator on.
| Têtu v1.0 | Têtu-2 instruct v2.0 | Têtu-2 chat v2.1 | |
|---|---|---|---|
| Multi-turn follow-up (1,475 turns) | 93.2 % | 71.0 % | 61.1 % |
| Concepts (350) | 75.4 % | 62.9 % | 52.6 % |
| Template math (215) | 79.1 % | 64.7 % | 64.7 % |
Têtu v1, which computes much more per token, remains better at following a conversation over several turns.
Base model: on the same held-out text, the validation loss is 2.680 for Têtu-2 base and 2.838 for Têtu v0.2 base (about 2B training tokens).
Limitations
- It produces false statements, especially on rare topics. Do not use it for medical, legal or financial information, or for any important decision.
- Its reasoning on new problems is weak: it can mix up units (60 km/h for 30 min → "1,800 km") or divide by the wrong number in an average.
- v2.1 sometimes confuses "I" and "you": asked « what did I ask you then? », it may answer as if it had asked.
- It was trained on little English conversation and is not a coding model.
Usage
The architecture is not supported by transformers; use the bundled inference code.
git clone https://huggingface.co/tetu-AI/Tetu-2 && cd Tetu-2
pip install -r requirements.txt
python chat.py # chat with v2.1; --cpu without a GPU; --no-calc to disable the calculator
Text completion with the base model:
import torch
from tetu2.generate import Generator, load
model, tok = load("base", device="cuda") # or device="cpu", dtype=torch.float32
gen = Generator(model, tok)
for piece, doubt in gen.stream("La capitale de la France est", max_new=40, temp=0):
print(piece, end="")
License
CC BY-NC 4.0 (non-commercial). The fine-tuning data includes material under a non-commercial license and synthetic conversations. Pre-training sources: FineWeb-Edu, FineWeb2-HQ and FineMath (ODC-By), Wikipedia (CC BY-SA). The chat teacher, Qwen2.5-7B-Instruct, is Apache 2.0.
Citation
@misc{tetu2_2026,
title = {Têtu-2: a small French language model trained from scratch},
author = {tetu-AI},
year = {2026},
url = {https://huggingface.co/tetu-AI/Tetu-2}
}
- Downloads last month
- 149


