Skip to content
Other

Zenotron-Mini-0.5-Multilingual

мультиязычная LLM с упором на логику

1 min readupdated

Это preview-релиз (v0.5) от Open Zeno-Zero Research Lab на базе Qwen3.6-35B-A3B. Модель прошла post-training методом Full LoRA (BF16) на 30 000 инструкциях, поровну на русском, английском, французском и корейском. Всего 35B параметров, около 3B активных, контекст — 16 384 токена.

Ключевые особенности

  • База Qwen3.6-35B-A3B: ~3B активных параметров при 35B общих — ставка на эффективный inference.
  • Full LoRA (BF16) и контекст 16 384 токена.
  • Обучение на 30 000 инструкциях: по 25% RU, EN, FR и KO.
  • Global-MGSM: 89.20% против 75.50% у базовой модели (+13.70 п.п.).
  • Cross-lingual прирост на Global-MGSM: FR +21.2%, KO +16.8%, EN +9.6%, RU +7.2%.
  • Belebele: 93.44% против 92.56% (+0.89 п.п.).

Пример

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
MODEL_ID = "open-zzrl/Zenotron-Mini-0.5-Multilingual"
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)
messages = [{"role": "user", "content": "Реши задачу по шагам: ..."}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt")
inputs = inputs.to(model.device)
ПодробнееGlobal-MGSM: 89.20% vs 75.50% у Qwen3.6-35B-A3B; Belebele: 93.44% vs 92.56%.
Веса в BF16; доступны FP8, FP4 и GGUF, включая i-quants для llama.cpp/Ollama.
Это экспериментальный preview v0.5: по данным авторов, добавление KO и FR размыло качество целевых языков. В v1.0 обещают сфокусироваться на RU и EN.
Примеры
Discuss Telegram 1 206

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 67 GB · ready files74 GB
  • FP8 / INT8recommendedweights 33 GB37 GB
  • 4-bitweights 19 GB21 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # open-zzrl/Zenotron-Mini-0.5-Multilingual
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download open-zzrl/Zenotron-Mini-0.5-Multilingual --local-dir ~/models/Zenotron-Mini-0.5-Multilingual
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Qwen3.6-35B-A3B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 1All on Hugging Face
Merges 4All on Hugging Face

About the model

open-zzrl/Zenotron-Mini-0.5-Multilingual
Author
open-zzrl
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

2061

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments

Other

HTR-500M

HTR-500M