Xlnk/sql

1---2library_name: transformers3license: other4license_name: lfm1.05license_link: LICENSE6language:7- en8- ar9- zh10- fr11- de12- ja13- ko14- es15pipeline_tag: text-generation16tags:17- liquid18- lfm2.519- edge20base_model: LiquidAI/LFM2.5-1.2B-Base21---22 23<div align="center">24  <img 25    src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/2b08LKpev0DNEk6DlnWkY.png" 26    alt="Liquid AI" 27    style="width: 100%; max-width: 100%; height: auto; display: inline-block; margin-bottom: 0.5em; margin-top: 0.5em;"28  />29  <div style="display: flex; justify-content: center; gap: 0.5em; margin-bottom: 1em;">30    <a href="https://playground.liquid.ai/"><strong>Try LFM</strong></a> • 31    <a href="https://docs.liquid.ai/lfm"><strong>Documentation</strong></a> • 32    <a href="https://leap.liquid.ai/"><strong>LEAP</strong></a>33  </div>34</div>35 36# LFM2.5-1.2B-Instruct37 38LFM2.5 is a new family of hybrid models designed for **on-device deployment**. It builds on the LFM2 architecture with extended pre-training and reinforcement learning.39 40- **Best-in-class performance**: A 1.2B model rivaling much larger models, bringing high-quality AI to your pocket.41- **Fast edge inference**: 239 tok/s decode on AMD CPU, 82 tok/s on mobile NPU. Runs under 1GB of memory with day-one support for llama.cpp, MLX, and vLLM.42- **Scaled training**: Extended pre-training from 10T to 28T tokens and large-scale multi-stage reinforcement learning.43 44![image](https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/dxnYF2fuLpulismtFSGFi.png)45 46Find more information about LFM2.5 in our [blog post](https://www.liquid.ai/blog/introducing-lfm2-5-the-next-generation-of-on-device-ai).47 48## 🗒️ Model Details49 50| Model | Parameters | Description |51|-------|------------|-------------|52| [LFM2.5-1.2B-Base](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Base) | 1.2B | Pre-trained base model for fine-tuning |53| [**LFM2.5-1.2B-Instruct**](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct) | 1.2B | General-purpose instruction-tuned model |54| [LFM2.5-1.2B-Thinking](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Thinking) | 1.2B | General-purpose reasoning model |55| [LFM2.5-1.2B-JP](https://huggingface.co/LiquidAI/LFM2.5-1.2B-JP) | 1.2B | Japanese-optimized chat model |56| [LFM2.5-VL-1.6B](https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B) | 1.6B | Vision-language model with fast inference |57| [LFM2.5-Audio-1.5B](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B) | 1.5B | Audio-language model for speech and text I/O |58 59LFM2.5-1.2B-Instruct is a general-purpose text-only model with the following features:60 61- **Number of parameters**: 1.17B62- **Number of layers**: 16 (10 double-gated LIV convolution blocks + 6 GQA blocks)63- **Training budget**: 28T tokens64- **Context length**: 32,768 tokens65- **Vocabulary size**: 65,53666- **Knowledge cutoff**: Mid-202467- **Languages**: English, Arabic, Chinese, French, German, Japanese, Korean, Spanish68- **Generation parameters**:69  - `temperature: 0.1`70  - `top_k: 50`71  - `top_p: 0.1`72  - `repetition_penalty: 1.05`73 74| Model | Description |75|-------|-------------|76| [**LFM2.5-1.2B-Instruct**](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct) | Original model checkpoint in native format. Best for fine-tuning or inference with Transformers and vLLM. |77| [LFM2.5-1.2B-Instruct-GGUF](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct-GGUF) | Quantized format for llama.cpp and compatible tools. Optimized for CPU inference and local deployment with reduced memory usage. |78| [LFM2.5-1.2B-Instruct-ONNX](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct-ONNX) | ONNX Runtime format for cross-platform deployment. Enables hardware-accelerated inference across diverse environments (cloud, edge, mobile). |79| [LFM2.5-1.2B-Instruct-MLX](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct-MLX-8bit) | MLX format for Apple Silicon. Optimized for fast inference on Mac devices using the MLX framework. |80 81We recommend using it for agentic tasks, data extraction, and RAG. It is not recommended for knowledge-intensive tasks and programming.82 83### Chat Template84 85LFM2.5 uses a ChatML-like format. See the [Chat Template documentation](https://docs.liquid.ai/lfm/key-concepts/chat-template) for details. Example:86 87```88<|startoftext|><|im_start|>system89You are a helpful assistant trained by Liquid AI.<|im_end|>90<|im_start|>user91What is C. elegans?<|im_end|>92<|im_start|>assistant93```94 95You can use [`tokenizer.apply_chat_template()`](https://huggingface.co/docs/transformers/en/chat_templating#using-applychattemplate) to format your messages automatically.96 97### Tool Use98 99LFM2.5 supports function calling as follows:100 1011. **Function definition**: We recommend providing the list of tools as a JSON object in the system prompt. You can also use the [`tokenizer.apply_chat_template()`](https://huggingface.co/docs/transformers/en/chat_extras#passing-tools) function with tools.1022. **Function call**: By default, LFM2.5 writes Pythonic function calls (a Python list between `<|tool_call_start|>` and `<|tool_call_end|>` special tokens), as the assistant answer. You can override this behavior by asking the model to output JSON function calls in the system prompt.1033. **Function execution**: The function call is executed, and the result is returned as a "tool" role.1044. **Final answer**: LFM2 interprets the outcome of the function call to address the original user prompt in plain text.105 106See the [Tool Use documentation](https://docs.liquid.ai/lfm/key-concepts/tool-use) for the full guide. Example:107 108```109<|startoftext|><|im_start|>system110List of tools: [{"name": "get_candidate_status", "description": "Retrieves the current status of a candidate in the recruitment process", "parameters": {"type": "object", "properties": {"candidate_id": {"type": "string", "description": "Unique identifier for the candidate"}}, "required": ["candidate_id"]}}]<|im_end|>111<|im_start|>user112What is the current status of candidate ID 12345?<|im_end|>113<|im_start|>assistant114<|tool_call_start|>[get_candidate_status(candidate_id="12345")]<|tool_call_end|>Checking the current status of candidate ID 12345.<|im_end|>115<|im_start|>tool116[{"candidate_id": "12345", "status": "Interview Scheduled", "position": "Clinical Research Associate", "date": "2023-11-20"}]<|im_end|>117<|im_start|>assistant118The candidate with ID 12345 is currently in the "Interview Scheduled" stage for the position of Clinical Research Associate, with an interview date set for 2023-11-20.<|im_end|>119```120 121## 🏃 Inference122 123LFM2.5 is supported by many inference frameworks. See the [Inference documentation](https://docs.liquid.ai/lfm/inference/transformers) for the full list.124 125| Name | Description | Docs | Notebook |126|------|-------------|------|:--------:|127| [Transformers](https://github.com/huggingface/transformers) | Simple inference with direct access to model internals. | <a href="https://docs.liquid.ai/lfm/inference/transformers">Link</a> | <a href="https://colab.research.google.com/drive/1_q3jQ6LtyiuPzFZv7Vw8xSfPU5FwkKZY?usp=sharing"><img src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/vlOyMEjwHa_b_LXysEu2E.png" width="110" alt="Colab link"></a> |128| [vLLM](https://github.com/vllm-project/vllm) | High-throughput production deployments with GPU. | <a href="https://docs.liquid.ai/lfm/inference/vllm">Link</a> | <a href="https://colab.research.google.com/drive/1VfyscuHP8A3we_YpnzuabYJzr5ju0Mit?usp=sharing"><img src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/vlOyMEjwHa_b_LXysEu2E.png" width="110" alt="Colab link"></a> |129| [llama.cpp](https://github.com/ggml-org/llama.cpp) | Cross-platform inference with CPU offloading. | <a href="https://docs.liquid.ai/lfm/inference/llama-cpp">Link</a> | <a href="https://colab.research.google.com/drive/1ohLl3w47OQZA4ELo46i5E4Z6oGWBAyo8?usp=sharing"><img src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/vlOyMEjwHa_b_LXysEu2E.png" width="110" alt="Colab link"></a> |130| [MLX](https://github.com/ml-explore/mlx) | Apple's machine learning framework optimized for Apple Silicon. | <a href="https://docs.liquid.ai/lfm/inference/mlx">Link</a> | — |131| [LM Studio](https://lmstudio.ai/) | Desktop application for running LLMs locally. | <a href="https://docs.liquid.ai/lfm/inference/lm-studio">Link</a> | — |132 133Here's a quick start example with Transformers:134 135```python136from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer137 138model_id = "LiquidAI/LFM2.5-1.2B-Instruct"139model = AutoModelForCausalLM.from_pretrained(140    model_id,141    device_map="auto",142    dtype="bfloat16",143#   attn_implementation="flash_attention_2" <- uncomment on compatible GPU144)145tokenizer = AutoTokenizer.from_pretrained(model_id)146streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)147 148prompt = "What is C. elegans?"149 150input_ids = tokenizer.apply_chat_template(151    [{"role": "user", "content": prompt}],152    add_generation_prompt=True,153    return_tensors="pt",154    tokenize=True,155).to(model.device)156 157output = model.generate(158    input_ids,159    do_sample=True,160    temperature=0.1,161    top_k=50,162    top_p=0.1,163    repetition_penalty=1.05,164    max_new_tokens=512,165    streamer=streamer,166)167```168 169## 🔧 Fine-Tuning170 171We recommend fine-tuning LFM2.5 for your specific use case to achieve the best results.172 173| Name | Description | Docs | Notebook |174|------|-------------|------|----------|175| CPT ([Unsloth](https://github.com/unslothai/unsloth)) | Continued Pre-Training using Unsloth for text completion. | <a href="https://docs.liquid.ai/lfm/fine-tuning/unsloth">Link</a> | <a href="https://colab.research.google.com/drive/10fm7eNMezs-DSn36mF7vAsNYlOsx9YZO?usp=sharing"><img src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/vlOyMEjwHa_b_LXysEu2E.png" width="110" alt="Colab link"></a> |176| CPT ([Unsloth](https://github.com/unslothai/unsloth)) | Continued Pre-Training using Unsloth for translation. | <a href="https://docs.liquid.ai/lfm/fine-tuning/unsloth">Link</a> | <a href="https://colab.research.google.com/drive/1gaP8yTle2_v35Um8Gpu9239fqbU7UgY8?usp=sharing"><img src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/vlOyMEjwHa_b_LXysEu2E.png" width="110" alt="Colab link"></a> |177| SFT ([Unsloth](https://github.com/unslothai/unsloth)) | Supervised Fine-Tuning with LoRA using Unsloth. | <a href="https://docs.liquid.ai/lfm/fine-tuning/unsloth">Link</a> | <a href="https://colab.research.google.com/drive/1vGRg4ksRj__6OLvXkHhvji_Pamv801Ss?usp=sharing"><img src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/vlOyMEjwHa_b_LXysEu2E.png" width="110" alt="Colab link"></a> |178| SFT ([TRL](https://github.com/huggingface/trl)) | Supervised Fine-Tuning with LoRA using TRL. | <a href="https://docs.liquid.ai/lfm/fine-tuning/trl">Link</a> | <a href="https://colab.research.google.com/drive/1j5Hk_SyBb2soUsuhU0eIEA9GwLNRnElF?usp=sharing"><img src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/vlOyMEjwHa_b_LXysEu2E.png" width="110" alt="Colab link"></a> |179| DPO ([TRL](https://github.com/huggingface/trl)) | Direct Preference Optimization with LoRA using TRL. | <a href="https://docs.liquid.ai/lfm/fine-tuning/trl">Link</a> | <a href="https://colab.research.google.com/drive/1MQdsPxFHeZweGsNx4RH7Ia8lG8PiGE1t?usp=sharing"><img src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/vlOyMEjwHa_b_LXysEu2E.png" width="110" alt="Colab link"></a> |180| GRPO ([Unsloth](https://github.com/unslothai/unsloth)) | GRPO with LoRA using Unsloth. | <a href="https://docs.liquid.ai/lfm/fine-tuning/unsloth">Link</a> | <a href="https://colab.research.google.com/drive/1mIikXFaGvcW4vXOZXLbVTxfBRw_XsXa5?usp=sharing"><img src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/vlOyMEjwHa_b_LXysEu2E.png" width="110" alt="Colab link"></a> |181 182## 📊 Performance183 184### Benchmarks185 186We compared LFM2.5-1.2B-Instruct with relevant sub-2B models on a diverse suite of benchmarks.187 188| Model | GPQA | MMLU-Pro | IFEval | IFBench | Multi-IF | AIME25 | BFCLv3 |189|-------|------|----------|--------|---------|----------|--------|--------|190| **LFM2.5-1.2B-Instruct** | 38.89 | 44.35 | 86.23 | 47.33 | 60.98 | 14.00 | 49.12 |191| Qwen3-1.7B (instruct)| 34.85 | 42.91 | 73.68 | 21.33 | 56.48 | 9.33 | 46.30 |192| Granite 4.0-1B | 24.24 | 33.53 | 79.61 | 21.00 | 43.65 | 3.33 | 52.43 |193| Llama 3.2 1B Instruct | 16.57 | 20.80 | 52.37 | 15.93 | 30.16 | 0.33 | 21.44 |194| Gemma 3 1B IT | 24.24 | 14.04 | 63.25 | 20.47 | 44.31 | 1.00 | 16.64 |195 196GPQA, MMLU-Pro, IFBench, and AIME25 follow [ArtificialAnalysis's methodology](https://artificialanalysis.ai/methodology/intelligence-benchmarking). For IFEval and Multi-IF, we report the average score across strict and loose prompt and instruction accuracies. For BFCLv3, we report the final weighted average score with a custom Liquid handler to support our tool use template. 197 198### Inference speed199 200LFM2.5-1.2B-Instruct offers extremely fast inference speed on CPUs with a low memory profile compared to similar-sized models.201 202![image](https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/dbbI-15p9re2ROhAkqnZm.png)203 204In addition, we are partnering with AMD, Qualcomm, and Nexa AI to bring the LFM2.5 family to NPUs. These optimized models are available through our partners, enabling highly efficient on-device inference. 205The following numbers have been calculated using 1K prefill and 100 decode tokens:206 207| Device                                               | Inference | Framework        | Model                | Prefill (tok/s) | Decode (tok/s) | Memory (GB) |208| ---------------------------------------------------- | --------- | ---------------- | -------------------- | --------------- | -------------- | ----------- |209| Qualcomm Snapdragon® X Elite                         | NPU       | NexaML           | LFM2.5-1.2B-Instruct | 2591            | 63             | 0.9GB       |210| Qualcomm Snapdragon® Gen4 (ROG Phone9 Pro)           | NPU       | NexaML           | LFM2.5-1.2B-Instruct | 4391            | 82             | 0.9GB       |211| Qualcomm Snapdragon® Gen4 (Samsung Galaxy S25 Ultra) | CPU       | llama.cpp (Q4_0) | LFM2.5-1.2B-Instruct | 335             | 70             | 719MB       |212| Qualcomm Snapdragon® Gen4 (Samsung Galaxy S25 Ultra) | CPU       | llama.cpp (Q4_0) | Qwen3-1.7B           | 181             | 40             | 1306MB      |213 214These capabilities unlock new deployment scenarios across various devices, including vehicles, mobile devices, laptops, IoT devices, and embedded systems.215 216## Contact217 218For enterprise solutions and edge deployment, contact [sales@liquid.ai](mailto:sales@liquid.ai).219 220## Citation221 222```bibtex223@article{liquidai2025lfm2,224  title={LFM2 Technical Report},225  author={Liquid AI},226  journal={arXiv preprint arXiv:2511.23404},227  year={2025}228}229```