--- license: other license_name: lfm-license-1.0 license_link: https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct/blob/main/LICENSE library_name: transformers pipeline_tag: text-generation language: - en - ar - zh - fr - de - ja - ko - es tags: - liquid - lfm2.5 - lfm2 - edge - on-device - tool-calling - heretic - uncensored - decensored - abliterated - reproducible - gguf - roleplay - local-llm base_model: LiquidAI/LFM2.5-1.2B-Instruct --- # LFM2.5-1.2B-Instruct-heretic
RACER IS OP

A decensored variant of [LiquidAI/LFM2.5-1.2B-Instruct](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct), produced with [Heretic](https://github.com/p-e-w/heretic) v1.4.0 (directional ablation / "abliteration"). Liquid's hybrid conv+attention architecture keeps its native Pythonic tool-call format, instruction-following, and multilingual coverage; refusal behaviour is suppressed via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning, so the base model's agentic ability is left largely intact. **Who this is for:** developers who want real instruction-following from a model that still fits on a phone. LFM2.5-1.2B-Instruct scores 86.2 on IFEval and 49.1 on BFCLv3 — genuinely agentic numbers — while decoding at 239 tok/s on an AMD CPU and 82 tok/s on a Snapdragon Gen4 NPU, in under 1 GB. That combination is rare: most models at this quality need a GPU. Use it for local agents, on-device assistants, RAG, and structured extraction where a 1.2B dense model would be too slow. Refusals drop from **98/100 to 2/100** — the lowest count in the collection. ## Runs anywhere | Device | Recommended quant | Weights | | :--- | :--- | :--- | | Phone / browser (4 GB RAM) | Q4_K_M | ~0.68 GB | | Raspberry Pi 5 (4 GB) | Q4_K_M | ~0.68 GB | | Laptop (8 GB, Apple Silicon) | Q8_0 | ~1.16 GB | | RTX 3060 / 4060 (8-12 GB) | Q6_K | ~0.90 GB | | Server / multi-tenant | Q5_K_M | ~0.79 GB | Weights only, at this model's 1.17B native size; Liquid measures the full runtime footprint at 0.9 GB on a Snapdragon X Elite NPU with 32K of context. The hybrid architecture spends 10 of its 16 layers in cheap convolution blocks, so this model is decode-bandwidth-bound: fast RAM and memory bandwidth matter far more than raw compute. OOM? Drop one quant level; the Q2_K tier (~0.45 GB) is there if you truly need it. ## Abliteration parameters Trial 135 of a 200-trial Heretic run (seed `3181001368`). | Parameter | Value | | :-------- | :---: | | **direction_index** | 10.60 | | **attn.o_proj.max_weight** | 1.42 | | **attn.o_proj.max_weight_position** | 9.08 | | **attn.o_proj.min_weight** | 1.26 | | **attn.o_proj.min_weight_distance** | 8.88 | | **mlp.down_proj.max_weight** | 1.25 | | **mlp.down_proj.max_weight_position** | 12.88 | | **mlp.down_proj.min_weight** | 1.10 | | **mlp.down_proj.min_weight_distance** | 7.17 | ## Performance | Metric | This model | Original model ([LiquidAI/LFM2.5-1.2B-Instruct](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct)) | | :----- | :--------: | :---------------------------: | | **KL divergence** | 0.0657 | 0 *(by definition)* | | **Refusals** | 2/100 | 98/100 | Refusals on the harmful evaluation set drop from 98/100 to 2/100 — the base model was almost fully refusal-tuned, and nearly all of that is gone. The KL divergence of 0.0657 is moderate, so expect a little more drift in phrasing than the tightest edits in this collection; the instruction-following and tool-call format survive intact. ## Why abliteration instead of fine-tuning Fine-tuning a "helpful" persona on top of RLHF'd refusals fights the base model's training and tends to degrade coherence. Abliteration instead finds and edits the specific weight directions responsible for refusal, leaving the rest of the network (and its capabilities) untouched. See the [Heretic repo](https://github.com/p-e-w/heretic) and the [original abliteration writeup](https://huggingface.co/blog/mlabonne/abliteration) for the mechanism. > Made with ❤️ by **RACER IS OP** — follow for more uncensored models ## Files ### Safetensors | File | Size | |---|---| | `model.safetensors` | 2.18 GB | BF16, 1.17B parameters. The `reproduce/` directory carries the full Heretic recipe — `config.toml`, `requirements.txt`, the Optuna study journal, and SHA-256 sums — so this exact model can be regenerated bit-for-bit. Reproduce it with `heretic --reproduce reproduce/reproduce.json`. ### GGUF quantizations Full quantization set (14 quants + F16) produced with [llama.cpp](https://github.com/ggml-org/llama.cpp). | File | Format | Size | |---|---|---| | `LFM2.5-1.2B-Instruct-heretic-F16.gguf` | GGUF F16 | 2.18 GB | | `LFM2.5-1.2B-Instruct-heretic-Q2_K.gguf` | GGUF Q2_K | 0.45 GB | | `LFM2.5-1.2B-Instruct-heretic-IQ3_S.gguf` | GGUF IQ3_S | 0.52 GB | | `LFM2.5-1.2B-Instruct-heretic-Q3_K_S.gguf` | GGUF Q3_K_S | 0.52 GB | | `LFM2.5-1.2B-Instruct-heretic-Q3_K_M.gguf` | GGUF Q3_K_M | 0.56 GB | | `LFM2.5-1.2B-Instruct-heretic-Q3_K_L.gguf` | GGUF Q3_K_L | 0.59 GB | | `LFM2.5-1.2B-Instruct-heretic-IQ4_XS.gguf` | GGUF IQ4_XS | 0.62 GB | | `LFM2.5-1.2B-Instruct-heretic-Q4_K_S.gguf` | GGUF Q4_K_S | 0.65 GB | | `LFM2.5-1.2B-Instruct-heretic-Q4_0.gguf` | GGUF Q4_0 | 0.65 GB | | `LFM2.5-1.2B-Instruct-heretic-Q4_1.gguf` | GGUF Q4_1 | 0.71 GB | | `LFM2.5-1.2B-Instruct-heretic-Q4_K_M.gguf` | GGUF Q4_K_M | 0.68 GB | | `LFM2.5-1.2B-Instruct-heretic-Q5_K_S.gguf` | GGUF Q5_K_S | 0.77 GB | | `LFM2.5-1.2B-Instruct-heretic-Q5_K_M.gguf` | GGUF Q5_K_M | 0.79 GB | | `LFM2.5-1.2B-Instruct-heretic-Q6_K.gguf` | GGUF Q6_K | 0.90 GB | | `LFM2.5-1.2B-Instruct-heretic-Q8_0.gguf` | GGUF Q8_0 | 1.16 GB | LFM2 dense architecture (`lfm2`) - loads natively in llama.cpp / LM Studio / Jan. Day-one upstream support, so this also works with MLX, vLLM, SGLang, and Nexa-optimized NPU builds. Run `llama serve -hf saidutta69/LFM2.5-1.2B-Instruct-heretic` to pull the default quant. ## Quickstart ```bash # llama.cpp llama serve -hf saidutta69/LFM2.5-1.2B-Instruct-heretic ``` ```python # transformers (requires transformers>=5.0.0) from transformers import AutoModelForCausalLM, AutoTokenizer model_name = "saidutta69/LFM2.5-1.2B-Instruct-heretic" model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", dtype="bfloat16") tokenizer = AutoTokenizer.from_pretrained(model_name) messages = [{"role": "user", "content": "A train leaves at 14:05 travelling 80 km/h. A car leaves at 14:15 travelling 120 km/h on the same track. When does the car catch the train? Show your work."}] inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device) out = model.generate(**inputs, do_sample=True, temperature=0.1, top_k=50, repetition_penalty=1.05, max_new_tokens=2048) print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)) ``` Also runnable via Ollama, LM Studio, Jan, vLLM, SGLang, and OpenVINO. ### Generation parameters Liquid's recommended settings - low temperature with a top-k cut: ```python model.generate(**inputs, temperature=0.1, top_k=50, repetition_penalty=1.05) ``` ### Chat template LFM2.5 uses a ChatML-like format. See the [Chat Template documentation](https://docs.liquid.ai/lfm/key-concepts/chat-template) for details. ``` <|startoftext|>system You are a helpful assistant trained by Liquid AI. user What is C. elegans? assistant ``` ### Tool use LFM2.5 writes Pythonic calls: a Python list between `<|tool_call_start|>` and `<|tool_call_end|>`. Pass tool definitions as a JSON list in the system prompt, or hand them to `tokenizer.apply_chat_template(..., tools=tools)`, then feed the result back under a `tool` role. To get JSON instead, just ask for it in the system prompt. ### Speculative decoding Attach [LFM2.5-1.2B-Instruct-DSpark](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct-DSpark) (a 296M drafter) for roughly 2.5x faster decoding on Apple Silicon via Metal, with identical outputs. ## Model details | | | |---|---| | **Architecture** | `Lfm2ForCausalLM` (hybrid conv + attention) | | **Parameters** | 1.17B | | **Layers / heads** | 16 layers (10 double-gated conv + 6 GQA), 32 attention heads, 8 KV heads | | **Hidden / intermediate** | 2048 / 12288 | | **Convolution cache** | L_cache = 3 | | **Position embedding** | RoPE, theta = 1,000,000 | | **Context length** | 128,000 (config `max_position_embeddings`); Liquid's tuned default is 32,768 | | **Vocab** | 65,536 | | **Precision** | bfloat16 | | **Languages** | English, Arabic, Chinese, French, German, Japanese, Korean, Spanish | | **Base model** | [LiquidAI/LFM2.5-1.2B-Instruct](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct) | ## Responsible use Refusal suppression is deliberate and works as intended: this model will comply with requests the base model would refuse, including some it shouldn't. There is no safety filtering layered on top. You are responsible for how you deploy it — don't put this behind an unmoderated public-facing endpoint serving third parties. It inherits LFM2.5's factual limitations; abliteration removes refusal directions, it doesn't add capability or judgment. ## License Inherits the [LFM License 1.0](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct/blob/main/LICENSE) from the base model. ## Related - [LiquidAI/LFM2.5-1.2B-Instruct](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct) — the base model - [LFM2.5 technical report](https://arxiv.org/abs/2511.23404) — architecture and benchmarks - [RACER IS OP — Heretic Models — full collection](https://huggingface.co/collections/saidutta69/racer-is-op-heretic-models)