experimentalmachines/LFM2.5-1.2B-Instruct-heretic

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Sep 7, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

experimentalmachines/LFM2.5-1.2B-Instruct-heretic is a 1.2 billion parameter language model derived from LiquidAI/LFM2.5-1.2B-Instruct, featuring a 32768-token context length. This model has undergone a directional ablation process using 'heretic' to minimize refusal rates while preserving original model fidelity. It is specifically designed as an 'uncensored' version, with the safety training of the base model intentionally removed. The model is also available in an ExecuTorch format, optimized for Arm CPUs with 8-bit dynamic per-token activations and 4-bit grouped weights.

Loading preview...

Model Overview

experimentalmachines/LFM2.5-1.2B-Instruct-heretic is a 1.2 billion parameter instruction-tuned language model based on LiquidAI/LFM2.5-1.2B-Instruct. Its primary distinction is the intentional removal of the base model's safety training, achieved through a process called 'heretic'. This method involves directional ablation of the refusal direction in the residual stream, using a TPE search to balance minimizing refusal rates with maintaining fidelity to the original model (low KL divergence).

Key Characteristics

  • Abliterated Safety Training: The model is explicitly designed to be 'uncensored' by removing the safety guardrails present in its base model.
  • High Context Length: Supports a substantial context window of 32,768 tokens.
  • Optimized for Edge Devices: Available in an ExecuTorch format (.pte file) specifically optimized for Arm CPUs (Apple Silicon, Snapdragon, Dimensity, Tensor, Graviton) via KleidiAI. This version uses 8-bit dynamic per-token activations and 4-bit grouped weights (8da4w).
  • Efficient Inference: The ExecuTorch build features an int8 embedding table, static KV cache, and fused SDPA, with a 2048-token prefill chunking mechanism.
  • Merged Weights: The ablation process results in merged weights, allowing the model to load as a standard Lfm2ForCausalLM without special handling.

Intended Use Cases

This model is suitable for applications where the explicit removal of safety training is desired or required, offering a less constrained response generation. Its ExecuTorch optimization makes it particularly well-suited for deployment on Arm-based edge devices and mobile platforms requiring efficient, quantized inference.

Finetunes
1 models