LFM3-1.7B Uncensored
A compact, community-modified text model for private chat on phones and computers. AtomicChat reduces refusal behavior in Liquid AI’s LFM3-1.7B and distributes GGUF files starting at 1.13 GB. Use OnDevice on compatible iPhones; Android loading remains unverified.
What do you want to do?
Choose a task. Get the software, files, settings, and steps to run it.
Set up private desktop chat with LFM3-1.7B Uncensored
Chat privately with LFM3-1.7B Uncensored in your own browser. Download the exact AtomicChat GGUF, run llama.cpp on Windows, macOS, or Linux, and keep conversations on your computer without a cloud account.
Set up private phone chat with LFM3-1.7B Uncensored
Set up private local chat with AtomicChat’s compact LFM3 model on your phone. Use OnDevice on iPhone, or follow the documented Android import steps with exact-model compatibility clearly marked as unverified.
Open a guide for its software, hardware requirements, and setup steps.
At a glance
- Parameters
- 1.73B · 1.7B model class
- Architecture
- LFM3 tri-hybrid · convolutions, sliding-window and global attention
- Context length
- 256K native · suggested phone starting point: 4,096 tokens
- License
- LFM Open License v1.0 · see the linked terms
- Software
- OnDevice: Local AI Studio, llama.cpp with LFM3 support
Using it on iPhone
Use this setup on a compatible iPhone with enough available memory for the model. The confirmed run used an iPhone 16; that is a tested example, not an exclusive device requirement. PocketPal failed to load the same file in that test. These observations apply to the reported setup; no speed, memory benchmark, or Android result was supplied.
Android instructions in the workflow are based on the apps’ current import documentation. Exact-model compatibility is still unverified. Downloading a GGUF is separate from successfully loading its architecture.
Choose your download
Use AD-Q4_K_M first. File sizes below are decimal GB, not the amount of RAM required to run the model.
Keep at least 3 GB free for the smaller file and a possible imported copy. This storage allowance is an estimate. Q6 and longer context need additional headroom.
What changed in the uncensored version?
AtomicChat reports these results for its modification of Liquid AI’s model. They are not independent YouRunAI benchmarks or phone performance measurements.
| Evaluation | Original | Uncensored |
|---|---|---|
| Refusals on harmful prompts | 61.7% | 0.0% |
| False refusals · XSTest | 5.2% | 0.0% |
| MMLU · 2,000 questions | 41.0% | 41.0% |
Zero refusals in a finite evaluation does not guarantee that every request will be answered. The unchanged MMLU result does not establish unchanged performance on all tasks.
Read AtomicChat’s evaluationBest for
- Keep lightweight text conversations on your phone using a local inference app.
- Chat privately in your browser on Windows, macOS, or Linux using llama.cpp.
- Experiment with short summaries, rewriting, and simple questions, checking the answers yourself.
Tradeoffs
- Reduced refusal behavior does not improve factual accuracy. Liquid AI cautions against complex multi-step reasoning and agentic coding with the small base model.
- The 256K native context ceiling is not a practical phone setting or a measured limit of this GGUF in OnDevice. Begin with 4,096 tokens where configurable and reduce it if memory is tight.
- GGUF import alone does not establish LFM3 support. PocketPal failed to load this file in the user’s test; that observation does not describe every version or device.
- Layla and ChatterUI document local GGUF importing, but loading this exact AtomicChat checkpoint on Android has not been verified for this guide.
- Offline generation requires a local model and disabled cloud/web features. App downloads, updates, and optional online tools still use a network connection.
Ways to run it
OnDevice: Local AI Studio · iPhone · OnDevice · iOS 18+
This route is for compatible iPhones with iOS 18 or later and enough available memory. OnDevice documents LFM3 support; performance and model loading depend on the device. Follow the linked phone workflow for model import and everyday offline chat. The app version, memory use, and token speed of the reported run were not recorded.
Official instructionsLayla / ChatterUI · Android · experimental import paths
These apps document importing a local GGUF; their packaged runtime support for this exact LFM3 checkpoint is unverified. Follow the Android section’s documented steps, then confirm model load and offline generation. An unsupported-architecture error needs a runtime update; a larger quantization will not fix it.
Official instructionsllama.cpp with LFM3 support · Desktop · publisher-documented runtime
AtomicChat requires a build containing LFM3 support (upstream PR #31252; its producer revision was 19cce69). Use the model card’s exact runtime instructions. No desktop application or hardware combination was independently tested for this page.
Official instructionsWhat to expect
Start with the 1.13 GB AD-Q4_K_M file in OnDevice on a compatible iPhone. Use the phone guide for iPhone and experimental Android imports, or the desktop guide for private browser chat with llama.cpp on Windows, macOS, and Linux.
Explore its uses
Sources and files
Primary sources and files behind this model page.
Keep exploring
Bonsai 2 · 27B
A compressed 27B-class reasoning model with publisher-provided GGUF packs and a dedicated llama.cpp fork for CUDA, Metal, and CPU.
DeepSeek V4.1 Flash
An open-weight model for reasoning, coding, agents, and image understanding. DeepSeek V4.1 Flash combines a 552B backbone with sparse activation and up to one million tokens of context. Choose the original checkpoint or a community uncensored variant below.
Qwen3.8
Explore 4 versions of Qwen3.8. Compare model sizes, packages, hardware requirements, and available setups.