Instructions to use lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S
# Run inference directly in the terminal:
llama cli -hf lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S
# Run inference directly in the terminal:
llama cli -hf lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S
# Run inference directly in the terminal:
./build/bin/llama cli -hf lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S

Use Docker

docker model run hf.co/lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S

Llama new
LM Studio
Jan

vLLM

How to use lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S

Ollama
How to use lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF with Ollama:
```
ollama run hf.co/lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S
```
Unsloth Desktop

How to use lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF with Docker Model Runner:
```
docker model run hf.co/lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S
```

Lemonade

How to use lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S

Run and chat with the model

lemonade run user.LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF-IQ2_S

List all available models

lemonade list

Hermes Agent

How to use lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF:IQ2_S" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

LFM2.5-24B-A2B — Heretic uncensor, GSQ-RCO-style GGUF

Per-tensor mixed-precision GGUF builds of the Heretic-abliterated LiquidAI/LFM2.5-24B-A2B, made for llama.cpp.

What this model is

LiquidAI/LFM2.5-24B-A2B is a ~24B mixture-of-experts text model in Liquid AI's lfm2_moe family, activating only ~2B parameters per token. The GGUF layout has 40 blocks: 38 MoE blocks with 64 experts each (1536×2048), 30 blocks with a short-convolution stack (shortconv.in_proj/out_proj/conv), and 10 blocks with full attention (2048×512 K/V). Embeddings are tied (there is no output.weight), the vocab is 128,000 tokens, and the context length is 131,072. Routed experts are 95.7% of all parameters.

License: LFM Open License v1.0 — local and research use is free; commercial use is allowed for entities under the $10M annual revenue threshold. Vendor card: https://huggingface.co/LiquidAI/LFM2.5-24B-A2B

Files

This is a text-only model: no MTP/NextN draft block, vision tower, or projector tensors are included, and no mmproj file is needed.

File	Size	BPW	Expert ladder (38 MoE blocks)
`quant/LFM2.5-24B-A2B-Heretic-GSQ-RCO-IQ2_XS.gguf`	6.99 GiB	2.50	17 edge @ IQ2_S, 1 @ IQ2_XS, 20 inner @ IQ2_XXS
`quant/LFM2.5-24B-A2B-Heretic-GSQ-RCO-IQ2_S.gguf`	7.56 GiB	2.70	9 edge @ IQ3_XXS, 2 @ IQ2_S, 27 inner @ IQ2_XS
`quant/LFM2.5-24B-A2B-Heretic-GSQ-RCO-IQ3_XXS.gguf`	8.39 GiB	3.01	25 edge @ IQ3_XXS, 2 @ IQ2_S, 11 inner @ IQ2_XS
`quant/LFM2.5-24B-A2B-Heretic-GSQ-RCO-IQ3_S.gguf`	9.78 GiB	3.50	32 edge @ IQ3_S, 3 @ IQ3_XXS, 3 inner @ IQ2_S

Only the final quantized builds are hosted here. The intermediate Q8_0 source and the importance matrix are linked below instead of being mirrored in this repository.

Source and provenance

The uncensor was produced with Heretic: a training-free abliteration pass that estimates the model's refusal direction, projects it out of selected weight matrices, and keeps trials under a KL bound with behavioural checks.

vendor BF16
  → Heretic abliteration kept at BF16 (OS-Software)
  → Q8_0 GGUF (effectively lossless)
  → these IQ2_XS / IQ2_S / IQ3_XXS / IQ3_S builds

BF16 uncensored master: OS-Software/LFM2.5-24B-A2B-Uncensored-Heretic
Q8_0 source: OS-Software/LFM2.5-24B-A2B-Uncensored-Heretic-GGUF → LFM2.5-24B-A2B-Uncensored-Heretic-Q8_0.gguf
Importance matrix: mradermacher/LFM2.5-24B-A2B-Uncensored-Heretic-i1-GGUF → LFM2.5-24B-A2B-Uncensored-Heretic.imatrix.gguf — coverage: 516 of 517 significant 2-D matrices. The only gap is the tied token_embd.weight, which is pinned at Q8_0 and needs no rows.

Quantization

"GSQ-RCO-style" means a format-compatible equivalent of the GSQ/RCO edge-weighted mixed-precision recipe; the published GSQ/RCO engines cannot emit GGUF.

The edge-weighted ladder runs on ffn_(gate|up|down)_exps; 95.7% of parameters live in routed experts, so the experts carry the compression. Fixed policies:

token_embd Q8_0 (tied, so it also serves the output header);
router ffn_gate_inp Q8_0;
attention: attn_q / attn_output Q6_K, attn_k / attn_v Q8_0;
shortconv: in_proj / out_proj Q6_K, the small 3×2048 conv Q8_0;
dense MLP blocks 0 and 1: Q6_K;
1-D norms and biases: fp32.

All four builds were smoke-tested with CPU-only llama.cpp tooling: they load, stay coherent, and produce reasoning output. The 8–11 t/s CPU number is a smoke signal, not a serving number.

Usage

llama-cli -m quant/LFM2.5-24B-A2B-Heretic-GSQ-RCO-IQ3_S.gguf

IQ3_S is the safer default; IQ3_XXS is the smaller option.

Honest limits

Heretic's behavioural test suite belongs to the community work, not this repository. These builds were checked for load, run, and coherence only.
The IQ2/IQ3 tiers inherit quantization loss. For strict quality comparisons, use OS-Software's Q6_K or BF16 path.
This llama.cpp build supports the lfm2_moe architecture out of the box.

Downloads last month: 1,970

GGUF

Model size

24B params

Architecture

lfm2moe

Hardware compatibility

2-bit

3-bit

Model tree for lavendercomputing/LFM2.5-24B-A2B-Heretic-GSQ-RCO-GGUF

Base model

LiquidAI/LFM2.5-24B-A2B-Base

Finetuned

LiquidAI/LFM2.5-24B-A2B

Quantized

(86)

this model