Skip to main content
Back

AtomicChat Strips LFM3-1.7B's Safety Limits Into a 1.13 GB Local File

A community fine-tune strips safety refusals from Liquid AI's tri-hybrid LFM3-1.7B and ships it as a 1.1GB GGUF that runs fully offline.

Model, Full Access
  • AtomicChat released a safety-ablated GGUF of Liquid AI's LFM3-1.7B, hitting 150K+ downloads.
  • The Q4_K_M build is 1.13 GB, Q6_K is 1.59 GB, both run offline.
  • Base model scores 49.37 across 29 benchmarks, beating Qwen3.5-9B on Liquid's own eval.
  • Supports 256K context and is post-trained for Hermes Agent, Pi, and OpenClaw harnesses.
  • Requires the latest llama.cpp build since LFM3 uses a custom tri-hybrid architecture.
  • Shipped under the LFM Open License v1.0, with a $10M revenue cap on free commercial use.

AtomicChat has published a community GGUF release of Liquid AI’s LFM3-1.7B that is designed to reduce built-in refusal behavior. The repository labels the model “uncensored” and offers two quantized files for local inference. Hugging Face showed more than 150,000 downloads during the release’s first weeks, though that count changes continuously.

The publisher describes the release as safety-ablated, meaning its modifications target alignment behavior that causes the assistant to decline some prompts. Its knowledge, architecture, context window, and reasoning limits still come from the upstream LFM3-1.7B model. The short model card provides no refusal-rate, quality-regression, or safety evaluation, so “uncensored” remains a repository label rather than a measured capability.

Two files, clear trade-offs

GGUF packages model weights and metadata in a format commonly used by llama.cpp and other local inference runtimes. Quantization reduces the numerical precision of those weights, lowering storage and memory requirements at the cost of some output fidelity.

Build Size Precision Best fit
AD-Q4_K_M 1.13 GB About 4-bit Lowest storage and memory use
AD-Q6_K 1.59 GB About 6-bit Higher weight fidelity with greater memory use

The repository says both files use importance-matrix calibration, which uses sampled activation data to preserve influential weights during quantization. Runtime memory will exceed the download size because inference also allocates buffers, model metadata, and a key-value cache. That cache grows with prompt length, so using the full 256K context can require far more memory than loading the 1.13 GB file alone.