norallm/normistral-11b-thinking-gguf overview

{
  "metadata": {},
  "card_data": {
    "license": "apache-2.0",
    "language": [
      "nb",
      "nn",
      "no"
    ],
    "base_model": [
      "norallm/normistral-11b-thinking"
    ],
    "library_name": "transformers",
    "pipeline_tag": "text-generation",
    "tags": [
      "norwegian",
      "bokmaal",
      "nynorsk"
    ],
    "frontmatter": {
      "license": "apache-2.0",
      "language": [
        "nb",
        "nn",
        "'no'"
      ],
      "base_model": [
        "norallm/normistral-11b-thinking"
      ],
      "library_name": "transformers",
      "pipeline_tag": "text-generation",
      "tags": [
        "norwegian",
        "bokmaal",
        "nynorsk"
      ]
    },
    "hero_image_url": "https://huggingface.co/norallm/normistral-11b-warm/resolve/main/images/puffin_2.png",
    "summary": "![](https://huggingface.co/norallm/normistral-11b-warm/resolve/main/images/puffin_2.png) This is our instruction-tuned NorMistral-11B language model for Norwegian, trained on open datasets and released under Apache 2.0 license. The model has undergone extensive fluency-preserving reinforcement learning according to our paper Fluent Alignment with Disfluent Judges: Post-training for Lower-resource Languages. The repository contains GGUF files with different amounts of quantization. We also provide a working .modelfile, which contains the official chat template converted to Go (as used by llama.cpp and ollama). The F16 model is also available for direct use at https://ollama.com/LTG/normistral-11b-thinking. **The model is freely available in our public chat interface: https://chat.llm.sigma2.no/** _____",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlicense: apache-2.0\nlanguage:\n- nb\n- nn\n- 'no'\nbase_model:\n- norallm/normistral-11b-thinking\nlibrary_name: transformers\npipeline_tag: text-generation\ntags:\n- norwegian\n- bokmaal\n- nynorsk\n---\n\n![](https://huggingface.co/norallm/normistral-11b-warm/resolve/main/images/puffin_2.png)\n\nThis is our instruction-tuned [NorMistral-11B](https://huggingface.co/norallm/normistral-11b-long) language model for Norwegian, trained on [open datasets](https://huggingface.co/datasets/norallm/normistral-11b-thinking-training) and released under Apache 2.0 license. The model has undergone extensive fluency-preserving reinforcement learning according to our paper [Fluent Alignment with Disfluent Judges: Post-training for Lower-resource Languages](https://arxiv.org/abs/2512.08777).\n\nThe repository contains GGUF files with different amounts of quantization. We also provide [a working `.modelfile`](https://huggingface.co/norallm/normistral-11b-thinking-gguf/blob/main/normistral-11b-thinking.modelfile), which contains the official chat template converted to Go (as used by `llama.cpp` and `ollama`). The F16 model is also available for direct use at https://ollama.com/LTG/normistral-11b-thinking.\n\n**The model is freely available in our public chat interface: https://chat.llm.sigma2.no/**\n\n_____\n\n## License\n\nWe release the model under Apache 2.0 license to indicate that we do not impose any additional constraints on the model weights.\nHowever, we do not own the data in the training collection.\n\n_____\n\n## Training and data\n\nGenerally speaking, the training follows our fluency-preserving post-training setup from [Fluent Alignment with Disfluent Judges: Post-training for Lower-resource Languages](https://arxiv.org/abs/2512.08777).\n\nThe training data is published alongside the model at [norallm/normistral-11b-thinking-training](https://huggingface.co/datasets/norallm/normistral-11b-thinking-training). Training code will be available at [github.com/ltgoslo/normistral-post-training](https://github.com/ltgoslo/normistral-post-training).\n\n**1. Supervised finetuning (SFT)**\n\nWe start by \"injecting\" the instruction-following and reasoning capabilities by SFT training on English responses and reasoning traces from [Kimi-K2-Thinking](https://huggingface.co/moonshotai/Kimi-K2-Thinking). The full SFT collection is published in [train_sft.jsonl](https://huggingface.co/datasets/norallm/normistral-11b-thinking-training/blob/main/train_sft.jsonl).\n\n**2. Reinforcement learning (d-RLAIF)**\n\nThe short SFT stage is followed by on-policy training on a large collection of Norwegian (Bokmål and Nynorsk) prompts (also available at [norallm/normistral-11b-thinking-training](https://huggingface.co/datasets/norallm/normistral-11b-thinking-training)). The specific setup of d-RLAIF (direct reinforcement learning from AI feedback) and its motivation is extensively described in [our paper](https://arxiv.org/abs/2512.08777). The \"AI\" reward model used here is [Mistral-Large-Instruct-2411](https://huggingface.co/mistralai/Mistral-Large-Instruct-2411).\n\n_____\n\n## Evaluation\n\nWe compared NorMistral against state-of-the-art instruction-tuned models of similar size. What follows is a preliminary evaluation on a generative version of [NorEval](https://arxiv.org/abs/2504.07749) (that is still work-in-progress). The responses from all evaluated models below are fully available for closer inspection at [norallm/normistral-11b-thinking-evaluation](https://huggingface.co/datasets/norallm/normistral-11b-thinking-evaluation).\n\n**Classification tasks**\n\nAll classification scores are reported as accuracy. NoReC sentiment analysis is done on sentence level. The generative scores (NorRewrite and Norsummarize) are reported as the average win-rates against `Llama-3.1-8B` evaluated using LLM-as-a-judge setup with `Llama-3.3-70B` (see [NorEval](https://arxiv.org/abs/2504.07749) for more information). * denotes \"thinking\" models.\n\n\n| Model            | NoReC_binary | NoReC_ternary | NorIdiom_NB | NorIdiom_NN | NorCSQA_NB | NorCSQA_NN |\n|------------------|--------------|---------------|:-----------:|:-----------:|------------|------------|\n| NorMistral-11B*  | **86.3**     | 65.2          | **55.7**    | **27.7**    | 70.7       | 64.2       |\n| Llama-3.1-8B     | 79.8         | 52.9          | 12.7        | 6.7         | 64.0       | 57.9       |\n| Mistral-Nemo-12B | 67.9         | 49.1          | 12.9        | 8.5         | 61.6       | 49.5       |\n| Qwen3-15B*       | 83.5         | **69.6**      | 22.1        | 13.2        | **83.8**   | 71.6       |\n| Gemma3-12B       | 85.2         | 67.1          | 43.7        | 23.7        | 81.9       | **80.0**   |\n| OLMo3-7B*        | 72.0         | 63.3          | 5.0         | 2.2         | 50.8       | 17.9       |\n| OLMo2-13B        | 32.8         | 13.2          | 3.5         | 2.2         | 48.0       | 45.3       |\n| Apertus-8B       | 78.4         | 58.8          | 34.3        | 15.7        | 69.2       | 63.2       |\n\n\n| Model            | NorOBQA_NB | NorOBQA_NN | NRK_NB   | NRK_NN   |NorRewrite   | NorSummarize  |\n|------------------|------------|------------|----------|----------|-------------|---------------|\n| NorMistral-11B*  | 83.0       | 84.4       | 58.8     | **62.3** |51.9         | 54.3          |\n| Llama-3.1-8B     | 78.5       | 71.1       | 49.8     | 46.2     |50.0         | 50.0          |\n| Mistral-Nemo-12B | 75.3       | 67.8       | 47.3     | 45.0     |42.5         | 39.2          |\n| Qwen3-15B*       | **94.4**   | **88.9**   | **63.3** | 55.9     |77.6         | **83.1**      |\n| Gemma3-12B       | 91.5       | **88.9**   | 59.8     | 58.4     |**86.8**     | 77.8.         |\n| OLMo3-7B*        | 70.5       | 54.4       | 43.3     | 35.9     |7.8          | 14.2          |\n| OLMo2-13B        | 55.3       | 56.7       | 45.3     | 39.4     |48.3         | 53.7          |\n| Apertus-8B       | 76.1       | 74.4       | 50.2     | 48.3     |39.6         | 42.1          |\n\n_____\n\n## Citation\n\n```bibtex\n@misc{samuel2025fluentalignmentdisfluentjudges,\n      title={Fluent Alignment with Disfluent Judges: Post-training for Lower-resource Languages}, \n      author={David Samuel and Lilja Øvrelid and Erik Velldal and Andrey Kutuzov},\n      year={2025},\n      eprint={2512.08777},\n      archivePrefix={arXiv},\n      primaryClass={cs.CL},\n      url={https://arxiv.org/abs/2512.08777}, \n}\n```\n\n```bibtex\n@inproceedings{samuel-etal-2025-small,\n    title = \"Small Languages, Big Models: {A} Study of Continual Training on Languages of {Norway}\",\n    author = \"Samuel, David  and\n      Mikhailov, Vladislav  and\n      Velldal, Erik  and\n      {\\O}vrelid, Lilja  and\n      Charpentier, Lucas Georges Gabriel  and\n      Kutuzov, Andrey  and\n      Oepen, Stephan\",\n    editor = \"Johansson, Richard  and\n      Stymne, Sara\",\n    booktitle = \"Proceedings of the Joint 25th Nordic Conference on Computational Linguistics and 11th Baltic Conference on Human Language Technologies (NoDaLiDa/Baltic-HLT 2025)\",\n    month = mar,\n    year = \"2025\",\n    address = \"Tallinn, Estonia\",\n    publisher = \"University of Tartu Library\",\n    url = \"https://aclanthology.org/2025.nodalida-1.61/\",\n    pages = \"573--608\",\n    ISBN = \"978-9908-53-109-0\",\n}\n```\n\n## Contact\n\nPlease write [a community message](https://huggingface.co/norallm/normistral-11b-thinking/discussions) or contact David Samuel (davisamu@ifi.uio.no) if you have any questions about this model.\n",
    "related_quantizations": []
  },
  "tags": [
    "transformers",
    "gguf",
    "norwegian",
    "bokmaal",
    "nynorsk",
    "text-generation",
    "nb",
    "nn",
    "no",
    "arxiv:2512.08777",
    "arxiv:2504.07749",
    "base_model:norallm/normistral-11b-thinking",
    "base_model:quantized:norallm/normistral-11b-thinking",
    "license:apache-2.0",
    "endpoints_compatible",
    "region:us",
    "conversational"
  ],
  "likes": 4,
  "downloads": 264,
  "gated": false,
  "private": false,
  "last_modified": "2025-12-25T20:29:09.000Z",
  "created_at": "2025-12-23T20:13:00.000Z",
  "pipeline_tag": "text-generation",
  "library_name": "transformers"
}

{
  "_id": "694af7cca8d391e3fb39e020",
  "id": "norallm/normistral-11b-thinking-gguf",
  "modelId": "norallm/normistral-11b-thinking-gguf",
  "sha": "10665e64e3fd7d111a4a9689238d99740b9fdf98",
  "createdAt": "2025-12-23T20:13:00.000Z",
  "lastModified": "2025-12-25T20:29:09.000Z",
  "author": "norallm",
  "downloads": 264,
  "likes": 4,
  "gated": false,
  "private": false,
  "pipeline_tag": "text-generation",
  "library_name": "transformers",
  "siblings_count": 10
}