Uraion Labs

Uraion Labs
Foundational systems research.

LFM3-1.7B — oQ4e (MLX)

Liquid AI's LFM3-1.7B, quantized for Apple Silicon with oMLX oQe mixed precision. Smallest and fastest. About 1 GB on disk.

Base model MLX oQ4e 256K context LFM Open License v1.0

LFM3-1.7B-MLX-oQ4e is an MLX quantization of LiquidAI/LFM3-1.7B, Liquid AI's tri-hybrid text model for local agents, tool calling and structured outputs. It was quantized and published by Uraion Labs with the oMLX oQ4e preset and runs natively on Apple Silicon Macs (M1 through M5) with oMLX and mlx-vlm >= 0.7.11.

  • Size: 1.04 GB of weights (~4.8 effective bits per weight, including group scales)
  • Speed: ~203 tokens/s decode on an M5 Pro (measured by Uraion Labs, see below)
  • Context: 256K tokens, native reasoning (<think>), XML tool calls, deferred tool search
  • Languages: 16, including English, Chinese, Arabic, Japanese, Korean, Hindi and the major European languages

Runtime note: LFM3 support in mlx-lm is still in progress upstream. Use oMLX or mlx-vlm 0.7.11+ (which loads LFM3 as a text-only model) today.

Pick your quant

Quant Base bits Mixed-precision profile Output head Size Decode speed* Best for Repository
oQ8e 8-bit Uniform 8-bit 8-bit 1.84 GB ~130 tok/s Maximum quality UraionLabs/LFM3-1.7B-MLX-oQ8e
oQ6e 6-bit Mixed 6/8-bit (33 layers @ 8-bit) 6-bit 1.45 GB ~152 tok/s Recommended: near-8-bit quality UraionLabs/LFM3-1.7B-MLX-oQ6e
oQ5e 5-bit Mixed 5/6/8-bit (29 @ 6b, 8 @ 8b) 6-bit 1.23 GB ~174 tok/s Faster, still strong quality UraionLabs/LFM3-1.7B-MLX-oQ5e
oQ4e (this repo) 4-bit Mixed 4/5/6/8-bit (37 @ 5b, 2 @ 6b, 4 @ 8b) 5-bit 1.04 GB ~203 tok/s Smallest and fastest (~1 GB) UraionLabs/LFM3-1.7B-MLX-oQ4e

*Single-stream decode speed measured by Uraion Labs on an Apple M5 Pro (64 GB), macOS 27, oMLX 0.7.3, 800 generated tokens, average of 2 runs. Your numbers will vary with chip, context length and settings.

All four variants answered a short math check correctly and produced a valid OpenAI-style get_weather tool call through the oMLX server (oQ4e and oQ8e tool-call tested). This is a smoke test, not a full benchmark run.

About this quantization

Property Value
Original model LiquidAI/LFM3-1.7B (Liquid AI)
Quantized by Uraion Labs
Quantization preset oQ4e (oMLX oQ, enhanced mixed precision)
Base weight precision 4-bit
Mixed-precision profile Mixed 4/5/6/8-bit (37 @ 5b, 2 @ 6b, 4 @ 8b)
Layer overrides 43 overrides: 37 boosted to 5-bit (attention q/k/v/o, FFN down-proj, embeddings), 2 to 6-bit, 4 attention gates to 8-bit
Embeddings / output head (tied) 5-bit
Group size / mode 64 / affine (per-group scale and bias)
Bit allocation Chosen per layer by oMLX oQ sensitivity analysis
Weight file 1.04 GB (1,037,627,479 bytes), single model.safetensors
Effective bits per weight ~4.80
Format MLX safetensors
Tested runtimes oMLX 0.7.3, mlx-vlm 0.7.11

Model information

Property Value
Architecture Lfm3ForCausalLM (lfm3), tri-hybrid
Parameters 1.73B
Layers 24: 13 gated short convolutions, 5 sliding-window attention (window 256), 6 global attention
Attention GQA, 16 query heads / 4 KV heads, head dim 128
Context length 262,144 tokens (input + output)
Vocabulary 128,000 tokens
Languages English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish
License LFM Open License v1.0

Quickstart

oMLX (recommended)

oMLX is an OpenAI-compatible server for Apple Silicon with continuous batching and SSD-backed KV cache. Download the model into your oMLX models folder (or use the model downloader in the oMLX app), then serve it:

hf download UraionLabs/LFM3-1.7B-MLX-oQ4e --local-dir ~/.omlx/models/LFM3-1.7B-MLX-oQ4e
omlx serve --model-dir ~/.omlx/models --port 8000

Call it from any OpenAI client, tools included:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="your-omlx-key")

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get the current weather for a city",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        },
    },
}]

response = client.chat.completions.create(
    model="LFM3-1.7B-MLX-oQ4e",
    messages=[{"role": "user", "content": "What's the weather in Paris right now?"}],
    tools=tools,
    temperature=0.1,
    extra_body={"top_k": 50, "repetition_penalty": 1.1},
)
print(response.choices[0].message.tool_calls)

mlx-vlm (CLI)

pip install -U "mlx-vlm>=0.7.11"

python -m mlx_vlm.generate \
  --model UraionLabs/LFM3-1.7B-MLX-oQ4e \
  --prompt "Explain how a heat pump works." \
  --max-tokens 1024 \
  --temperature 0.1 \
  --top-k 50 \
  --repetition-penalty 1.1

mlx-vlm (Python)

from mlx_vlm import load, generate

model, processor = load("UraionLabs/LFM3-1.7B-MLX-oQ4e")

messages = [{"role": "user", "content": "Explain how a heat pump works."}]
prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

result = generate(model, processor, prompt, max_tokens=1024, temperature=0.1, verbose=True)
print(result.text)

Agent harnesses

LFM3 was post-trained inside agent harnesses such as Hermes Agent, Pi and OpenClaw. Point any OpenAI-compatible harness at the oMLX endpoint above. See Liquid AI's Agent Harnesses guide.

Recommended sampling

From Liquid AI's official model card:

Use case Settings
General (chat, tool use, agents, coding) temperature=0.1, top_k=50, repetition_penalty=1.1
Math and STEM temperature=1.0, top_p=0.95, top_k=20

The bundled generation_config.json already uses the general settings.

Chat format, reasoning and tools

  • ChatML template (<|im_start|> / <|im_end|>). Use apply_chat_template(..., add_generation_prompt=True).
  • Reasoning is on by default. Pass enable_thinking=False to the template to turn it off (upstream notes this can be inconsistent on reasoning-heavy tasks). Earlier-turn reasoning is dropped unless preserve_thinking=True.
  • Tool calls are emitted as XML (<tool_call><function=name><parameter=...>). oMLX converts them to OpenAI tool_calls automatically.
  • Deferred tool search: mark tools with defer_loading=True and return tool_reference parts from your search tool. Details in Liquid AI's tool-use docs.

Upstream benchmarks

Reported by Liquid AI for the original BF16 LFM3-1.7B (vLLM, full context). Uraion Labs has not re-run these on the quantized weights. Lower-bit quants may score somewhat lower.

Benchmark LFM3-1.7B LFM2.5-2.6B MiniCPM5-2B Gemma 4 E4B Qwen3.5-4B Qwen3.5-9B
LiveClaw 45.60 14.70 30.37 20.36 19.94 35.77
Claw-Eval General (EN), average 68.64 62.63 60.74 56.14 61.96 66.53
BFCL V4 71.39 55.89 62.65 48.25 55.12 62.83
ToolSandbox 80.36 78.04 68.16 69.27 75.40 76.41
τ³-bench (4 domains), average 68.01 56.77 69.44 33.53 65.98 67.11
MCP-Atlas 32.57 16.45 26.38 6.19 41.37 42.67
SWE-bench Verified (OpenCode) 43.40 18.00 34.80 0.50 41.20 51.20
LiveCodeBench v6 73.46 60.57 70.52 71.00 61.99 69.86
IFEval 96.66 95.39 87.83 90.02 92.52 92.08
IFBench 73.03 60.48 62.08 44.56 38.16 48.84
IFStruct 93.72 82.18 72.65 78.45 87.05 89.20
GPQA Diamond 63.37 57.78 65.15 57.88 75.76 80.40
AA-Omniscience non-hallucination (public) 92.93 53.69 84.44 72.09 13.64 52.59
AA-LCR 52.83 6.00 49.34 39.33 52.67 62.00
RULER 128K 65.56 1.15 37.96 18.82 84.63 87.11
Average 49.37 31.55 39.19 31.74 39.27 47.24

Full table, settings and per-benchmark details: LiquidAI/LFM3-1.7B.

Limitations

  • Liquid AI positions LFM3-1.7B for local agents, tool calling, structured outputs and fine-tuning. It is not recommended for agentic coding or complex multi-step reasoning because of its size. Pair it with retrieval or search for knowledge-heavy tasks.
  • Quantization reduces precision. The 4-bit and 5-bit variants trade some accuracy for size and speed; use oQ6e or oQ8e when quality matters most.
  • Like any language model, it can produce inaccurate or biased output. Do not treat answers on legal, medical or financial topics as professional advice. Provided "AS IS", without warranty.

Attribution and modifications

LFM3-1.7B was created by Liquid AI. The architecture, training, benchmarks and capabilities described here come from the upstream release. All credit for the model belongs to Liquid AI.

Changes made by Uraion Labs: the weights in model.safetensors were quantized to MLX format, and config.json was updated with quantization metadata. The tokenizer, chat template, generation config and custom code files are unchanged from upstream. No retraining or fine-tuning was performed.

Upstream resources: model card · LFM3 blog post · docs · LFM3 collection · Discord

License

Distributed under the LFM Open License v1.0 (copy included in LICENSE), the same license as the original model. In short: free for research and non-commercial use, and for commercial use by organizations under US$10M in annual revenue. Commercial use above that threshold is not covered; contact Liquid AI. Read the full license before use.

Citation

@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},
}

Quantized and published by Uraion Labs. Foundational systems research.

Downloads last month
335
Safetensors
Model size
2B params
Tensor type
U32
·
BF16
·
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for UraionLabs/LFM3-1.7B-MLX-oQ4e

Quantized
(164)
this model

Collection including UraionLabs/LFM3-1.7B-MLX-oQ4e