Model record
LFM3-1.7B
Liquid AI · LFM Open License v1.0 · released · verified
What this model is
Liquid AI reports 1.73B parameters in a tri-hybrid architecture with 13 gated short-convolution layers, five sliding-attention layers, and six global-attention layers.
The architecture is dense. Every parameter is part of the active path, so memory and speed scale with the full model rather than a smaller expert subset.
Memory and quantization
Decimal GB from Liquid AI's official GGUF repository. QAD-Q4_0 is the quantization-aware distillation build; the other entries are the published K-quant and Q8 files.
| Quantization | Published files | 8K estimate | Smallest comfortable GPU | Smallest comfortable Mac | Context at smallest GPU / Mac | Speed estimate | Run commands |
|---|---|---|---|---|---|---|---|
| QAD-Q4_0 | 1.05 GB | 2.65 GB | NVIDIA GeForce RTX 5060 Ti · 16 GB · 16 GB VRAM | Mac mini M6 · 24 GB · 24 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | ollama run lfm3:1.7b-qad-q4_0 lms get "LFM3-1.7B"@qad-q4_0 |
| Q4_K_M | 1.10 GB | 2.70 GB | NVIDIA GeForce RTX 5060 Ti · 16 GB · 16 GB VRAM | Mac mini M6 · 24 GB · 24 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | ollama run lfm3:1.7b-q4_K_M lms get "LFM3-1.7B"@q4_k_m |
| Q5_K_M | 1.26 GB | 2.86 GB | NVIDIA GeForce RTX 5060 Ti · 16 GB · 16 GB VRAM | Mac mini M6 · 24 GB · 24 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | lms get "LFM3-1.7B"@q5_k_m |
| Q6_K | 1.43 GB | 3.03 GB | NVIDIA GeForce RTX 5060 Ti · 16 GB · 16 GB VRAM | Mac mini M6 · 24 GB · 24 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | lms get "LFM3-1.7B"@q6_k |
| Q8_0 | 1.85 GB | 3.45 GB | NVIDIA GeForce RTX 5060 Ti · 16 GB · 16 GB VRAM | Mac mini M6 · 24 GB · 24 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | ollama run lfm3:1.7b-q8_0 lms get "LFM3-1.7B"@q8_0 |
Estimate formula: published weight files + 0.1007 GB per 8K tokens × context / 8K + 1.5 GB runtime overhead. Sizes are decimal GB, not GiB. At 8K context, the smallest listed build is about 2.7 GB before OS and application reserve.
Conservative FP16 estimate from the six global-attention layers. The five sliding-attention layers use a 256-token window, and convolution state is excluded from the per-token linear estimate.
Ollama tags were checked against the linked Ollama library pages on 2026-11-02. LM Studio's lms get suffix selects a file when the search results expose that exact quantization; confirm the local key with lms ls.
KV method source: https://huggingface.co/LiquidAI/LFM3-1.7B/raw/main/config.json
Speed heuristic
Not benchmarked here. A 1.73B dense active path is generally fast on current CPUs and GPUs, but tri-hybrid kernel support still matters.
Active path: 1.73B parameters. This is a relative speed signal only; it is not a tokens/sec benchmark.
Runtimes
- Ollama
- llama.cpp
- MLX
- vLLM
ollama run lfm3:1.7blms get "LFM3-1.7B"@qad-q4_0
lms load <model-key-from-lms-ls> --gpu max --context-length 8192License restrictions
The LFM Open License v1.0 governs commercial use, redistribution, and acceptable-use terms. Read the linked license before embedding the model in a product.
This is a plain-language summary, not legal advice. Read the linked license before commercial deployment or redistribution.
Known issues
- Liquid AI positions this as a small agent and structured-output model, not a replacement for a larger coding or complex-reasoning model.
- The 256K context is a capacity limit; long agent traces still consume KV cache and can lower throughput.
- The official GGUF card requires Ollama v0.40.12 or newer for the LFM3 architecture.
FAQ
Model-specific questions
What is the smallest local file for LFM3-1.7B?
The smallest published weight package in this record is QAD-Q4_0 at 1.05 GB. The 8K runtime estimate adds 0.10 GB of KV cache and 1.5 GB overhead, for about 2.7 GB before operating-system reserve.
How much memory do the published quantizations need at 8K?
QAD-Q4_0 about 2.7 GB at 8K; Q4_K_M about 2.7 GB at 8K; Q5_K_M about 2.9 GB at 8K; Q6_K about 3.0 GB at 8K; Q8_0 about 3.4 GB at 8K. These are calculated estimates using the published weight sizes and the model's KV-cache assumption, not live memory measurements.
Which GPU and Mac configurations are comfortable for LFM3-1.7B?
Against the current catalog, the smallest configurations that clear the 20% headroom target are: QAD-Q4_0: NVIDIA GeForce RTX 5060 Ti · 16 GB; Q4_K_M: NVIDIA GeForce RTX 5060 Ti · 16 GB; Q5_K_M: NVIDIA GeForce RTX 5060 Ti · 16 GB; Q6_K: NVIDIA GeForce RTX 5060 Ti · 16 GB; Q8_0: NVIDIA GeForce RTX 5060 Ti · 16 GB. A larger configuration may still be preferable for longer context or faster memory bandwidth.
What is the right Ollama or LM Studio command for LFM3-1.7B?
Use the command attached to the specific quantization in the table. Ollama tags are taken from the checked Ollama library pages. For LM Studio, use lms get with the displayed quantization suffix, then load the resulting local key from lms ls with --gpu max and the desired context length.
What should I run instead if LFM3-1.7B does not fit?
This is already one of the smaller records in the catalog. Reduce context, choose the lowest published quantization, or compare the entry in the VRAM tier guides before moving to a larger machine.
Sources
- https://huggingface.co/LiquidAI/LFM3-1.7B
- https://www.liquid.ai/blog/lfm3
- https://huggingface.co/LiquidAI/LFM3-1.7B-GGUF
- https://huggingface.co/LiquidAI/LFM3-1.7B-MLX-4bit
- https://ollama.com/library/lfm3
- https://huggingface.co/LiquidAI/LFM3-1.7B/raw/main/config.json
This page is an independent compatibility guide, not official documentation. Verify the primary source before downloading or deploying a model.