LiquidAI/LFM3-1.7B · Hugging Face

Source: https://huggingface.co/LiquidAI/LFM3-1.7B

LFM3-1.7B is atext model for local agents and specialized applications, based on our new tri-hybrid architecture.

• **Agentic training:**post-trained for agentic harnesses such as Hermes Agent, Pi, and OpenClaw, with a 256K-token context window.
• **Designed for the edge:**at 4-bit precision, it takes about 1 GB of memory, so it fits on any phone or laptop.
• **Made to be specialized:**adapt the model to your domains, formats, and tools via quick single-GPU fine-tuning.

Find more information about the LFM3 family inour release blog post.

https://huggingface.co/LiquidAI/LFM3-1.7B#model-detailsModel details

TheLFM3 familyincludes base, text, vision, audio, omni, and decision models.

PropertySpecificationCheckpointPost-trained, text-only language modelParameters1.73BArchitecture24 layers: 13 gated short convolutions, 5 sliding-window attention, 6 global attentionContext length256K tokens (input + output)Vocabulary128,000 tokensLanguagesEnglish, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, PolishLicenseLFM Open License v1.0LFM3-1.7B is designed for local agents, tool calling, structured outputs, and domain-specific fine-tuning. Use retrieval or search for knowledge-heavy tasks. It is not recommended for agentic coding or complex multi-step reasoning due to its small size.

https://huggingface.co/LiquidAI/LFM3-1.7B#inferenceInference

Choose a checkpoint variant for your inference framework and hardware.

CheckpointFrameworkDescriptionBF16vLLM,SGLang,TransformersBF16 weights for GPU inference.FP8SGLangFP8 quantization for GPU inference.NVFP4SGLang4-bit floating-point quantization for GPU inference.GGUFllama.cpp,Ollama,LM StudioLocal inference on CPUs and GPUs.MLXMLX4-bit quantization for Apple Silicon.ONNXONNX RuntimeCross-platform inference with ONNX Runtime.Sampling parameters:

• General(chat, tool use, agents, coding):temperature=0\.1,top\_k=50, repetition\_penalty=1\.1
• Math & STEM(math, science):temperature=1\.0, top\_p=0\.95, top\_k=20

Transformers quick start:

from transformers import pipeline

generator = pipeline(
    "text-generation",
    model="LiquidAI/LFM3-1.7B",
    trust_remote_code=True,
)

messages = [{"role": "user", "content": "What is C. elegans?"}]
output = generator(
    messages,
    do_sample=True,
    temperature=0.1,
    top_k=50,
    repetition_penalty=1.1,
    max_new_tokens=2048,
)
print(output[0]["generated_text"][-1]["content"])

For local agent setup, see theAgent Harnesses guide.

Chat templateLFM3 uses this ChatML format, generated from the checkpoint’schat template:

<|startoftext|><|im_start|>system
You are a helpful assistant trained by Liquid AI.<|im_end|>
<|im_start|>user
Explain how a heat pump works.<|im_end|>
<|im_start|>assistant
<think>

Usetokenizer\.apply\_chat\_template\(\)withadd\_generation\_prompt=Trueto format messages automatically.

Reasoning is enabled by default. You can setenable\_thinking=Falseto disable thinking traces, but this behavior can be inconsistent, especially for reasoning-heavy tasks.

Reasoning from turns before the latest user message is removed unlesspreserve\_thinking=True.

Tool callingPass function schemas throughtools=\.\.\.toapply\_chat\_template\(\). The template formats tool definitions and results automatically. Generated calls use XML:

<tool_call>
<function=get_weather>
<parameter=city>
Paris
</parameter>
</function>
</tool_call>

Execute the call in your application. Add the parsed assistant call to history as atool\_callsentry, then append the result as atoolmessage. Keep passing the tool schemas on subsequent requests and preserve the assistant’s content and reasoning from the current tool-use turn.

function\.argumentsmust be a Python mapping. Decode JSON strings withjson\.loads\(\)before applying the template.

Tool searchLFM3 supports deferred tools. Setdefer\_loading=Trueto keep tool definitions out of the prompt until your harness’s search tool returns their names astool\_referenceparts. The chat template then inserts their definitions into<additional\_tools\>.

tools = [
    search_tool,
    {"type": "function", "defer_loading": True, "function": get_weather_schema},
]
messages.append({
    "role": "tool",
    "tool_call_id": call_id,
    "content": [
        {"type": "text", "text": "Found 1 tool."},
        {"type": "tool_reference", "name": "get_weather"},
    ],
})

Your harness implements and runs the search. Keep the fulltoolslist, including deferred tools, on every request. In the system prompt, tell the model when to search:

You only have one tool right now: search_tools, which searches a catalog of many more tools. If none of your current tools can handle a request, call search_tools to find one, then call the tool it returns.

Works with vLLM, SGLang, and LFM3-compatible llama.cpp builds. For llama.cpp, use a GGUF with the current chat template or load it with\-\-chat\-template\-file.

SeeDeferred tools and tool searchfor the full guide.

https://huggingface.co/LiquidAI/LFM3-1.7B#fine-tuningFine-tuning

LFM3-1.7B is designed foreasy customizationfor a given domain, task, or application viaTRL.

TrainingUse caseNotebookSFTExtract structured JSON from messy documentsDPOAnswer from retrieved documents, or declineGRPOCall the right tools in a 108-tool phone appCPTLearn a new language (Turkish) from raw text## https://huggingface.co/LiquidAI/LFM3-1.7B#performancePerformance

https://huggingface.co/LiquidAI/LFM3-1.7B#benchmarksBenchmarks

LFM3-1.7B can power agentic applications that used to need 9B parameters.

BenchmarkLFM3 1.7BLFM2.5 2.6BMiniCPM5 2BGemma 4 E4BQwen3.5 4BQwen3.5 9BAgentsLiveClaw45.6014.7030.3720.3619.9435.77Claw-Eval General (EN), average68.6462.6360.7456.1461.9666.53Claw-Eval General (EN), pass^351.4037.5234.5822.4342.9951.87BrowseComp-Plus (OpenClaw subset)37.0026.3316.8321.0024.6738.00WebArena-Verified38.423.029.1117.0621.9228.94EnterpriseOps-Gym28.207.9413.419.8625.9628.97AutomationBench (public)41.290.944.711.928.6715.10Tool useBFCL V471.3955.8962.6548.2555.1262.83ToolSandbox80.3678.0468.1669.2775.4076.41MCP-Atlas32.5716.4526.386.1941.3742.67τ³-bench (4 domains), average68.0156.7769.4433.5365.9867.11Agentic codingSWE-bench Verified (OpenCode)43.4018.0034.800.5041.2051.20SWE-bench Multilingual (OpenCode)23.008.0015.670.6723.3335.00Terminal-Bench 2.1 (OpenCode)12.920.005.620.0013.4817.98CodingLiveCodeBench v673.4660.5770.5271.0061.9969.86LiveCodeBench Pro 25Q231.7419.1629.3429.9424.2534.43SciCode28.4716.3227.7825.0023.0231.60Instruction followingIFEval96.6695.3987.8390.0292.5292.08IFBench73.0360.4862.0844.5638.1648.84IFStruct93.7282.1872.6578.4587.0589.20Multi-IF89.1084.8073.1981.0380.3284.44KnowledgeGPQA Diamond63.3757.7865.1557.8875.7680.40Humanity’s Last Exam (no tools)7.175.888.204.7712.8416.41Humanity’s Last Exam (with tools)13.909.6011.506.0011.6017.50AA-Omniscience (public)0.50-33.17-5.00-15.33-57.17-28.00AA-Omniscience accuracy (public)7.078.759.139.8315.6713.17AA-Omniscience non-hallucination (public)92.9353.6984.4472.0913.6452.59Long contextAA-LCR52.836.0049.3439.3352.6762.00RULER 128K65.561.1537.9618.8284.6387.11Average49.3731.5539.1931.7439.2747.24 Evaluation detailsAll models were served with vLLM in BF16 at their full context length: 262,144 tokens for LFM3-1.7B, Qwen3.5-4B, and Qwen3.5-9B, and 131,072 tokens for LFM2.5-2.6B, MiniCPM5-2B, and Gemma 4 E4B. Unless listed below, every eval uses temperature 0.6 and top_p 0.95, with no output limit other than the context length. The Average row is the unweighted mean of every score in the table.

BFCL V4: greedy decoding, up to 32,768 output tokens. ToolSandbox: greedy decoding, up to 1,024 tokens per turn. LiveClaw: OpenClaw agent, up to 8,192 tokens per turn. Claw-Eval: greedy decoding, 3 trials per task. BrowseComp-Plus: OpenClaw agent, greedy decoding, up to 8,192 tokens per turn. WebArena-Verified: up to 40 steps. MCP-Atlas: up to 256 turns. EnterpriseOps-GymandAutomationBench: up to 50 turns. SWE-bench: OpenCode agent, up to 250 turns. Terminal-Bench 2.1: OpenCode agent, up to 100 turns.

https://huggingface.co/LiquidAI/LFM3-1.7B#edge-inferenceEdge inference

At 32K context, LFM3-1.7B prefills at least 2.5x faster and decodes at least 1.8x faster than every compared model on all three devices, while using the least memory.

Evaluation detailsAll models run as Q4_0 GGUF files in llama.cpp with 32,768 tokens of context, batch size 1, 100 output tokens, and GPU offload. Prefill is the prompt processing speed and decode is the generation speed, both in tokens per second. LFM3-1.7B uses a llama.cpp build with LFM3 support, and the other models use llama.cpp release b11744.

AMD Ryzen AI Max+ 395: Radeon 8060S integrated GPU, Windows, Vulkan backend. Apple M5 Max: macOS, Metal backend. iPhone 17 Pro: Apple A19 Pro, iOS app built on llama.cpp. Memory: peak usage in GB (109bytes). On AMD, it adds the host and GPU peaks from the same run, which can slightly overstate usage if the two peaks do not coincide. On Apple devices, the process footprint already includes Metal allocations.

https://huggingface.co/LiquidAI/LFM3-1.7B#gpu-inferenceGPU inference

At 1,024 concurrent requests, LFM3-1.7B can produce 3.6 billion tokens/day on one NVIDIA B300.

https://huggingface.co/LiquidAI/LFM3-1.7B#trainingTraining

**Pre-training:**two stages build LFM3-1.7B-Base on 22T tokens in total.

• **Main pre-training:**compared with LFM2, the mix adds more scientific papers, repository-level code, math, and a wider range of knowledge sources.
• **Mid-training:**the last 3.5T tokens use curated, higher-quality data with a larger share of agentic traces and tool calls. This stage also extends the context window to 256K tokens.

**Post-training:**four stages develop conversation, reasoning, and tool-use capabilities.

• **Supervised fine-tuning (SFT):**1.75T tokens across various tasks, including chat, reasoning, tool calling, and agentic traces.
• **Mixed reinforcement learning:**many domains trained at once, from math and code to SQL and structured outputs. A length penalty in the reward keeps answers concise.
• **Multi-domain on-policy distillation (MOPD):**six specialist RL teachers (code, math, tool use, long context, agentic work, and enterprise workflows) are distilled back into a single model.
• **Agentic RL:**tasks from web browsing to office work are run inside real agent harnesses, preparing the model for deployment.

https://huggingface.co/LiquidAI/LFM3-1.7B#citationCitation

@article{liquidAI2026LFM3,
  author  = {Liquid AI},
  title   = {Introducing LFM3: Every Modality for Specialized Intelligence},
  journal = {Liquid AI Blog},
  year    = {2026},
  note    = {www.liquid.ai/blog/lfm3},
}

For questions and feedback, joinDiscord. For custom deployment and fine-tuning support,contact Liquid AI.

Downloads last month6,311

Model tree forLiquidAI/LFM3-1.7Bhttps://huggingface.co/docs/hub/model-cards#specifying-a-base-model

Spaces usingLiquidAI/LFM3-1.7B3

Collection includingLiquidAI/LFM3-1.7B

Article mentioningLiquidAI/LFM3-1.7B