Instructions to use IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-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 IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-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 IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF
# Run inference directly in the terminal:
llama cli -hf IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF
# Run inference directly in the terminal:
llama cli -hf IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF

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 IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF
# Run inference directly in the terminal:
./build/bin/llama cli -hf IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF

Use Docker

docker model run hf.co/IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF

Llama new
LM Studio
Jan

vLLM

How to use IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-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": "IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF

Ollama
How to use IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF with Ollama:
```
ollama run hf.co/IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF
```
Unsloth Desktop

How to use IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF

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": "IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF with Docker Model Runner:
```
docker model run hf.co/IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF
```

Lemonade

How to use IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF

Run and chat with the model

lemonade run user.LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-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 IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF

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 IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF

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 "IsValorum/LFM2.5-24B-A2B-APEX-I-NanoPlus-GGUF" \
  --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 APEX-I NanoPlus GGUF

Independently computed using dedicated cloud compute clusters, funded 100% out-of-pocket.

If you consider these releases to be of high quality and they help you in your day-to-day workflow, consider supporting on Ko-fi. Every contribution directly sustains future cloud compute hours and helps support my university studies.

The 24B Frontier Mixture-of-Experts On-Device Reasoning Specialist (2B Active Parameters) · Native 128K Context · Streamlined Ultra-Lean Footprint

Official APEX-I-NanoPlus quantization of LiquidAI/LFM2.5-24B-A2B. Calibrated with the high-coverage importance matrix (imatrix.gguf) generated by m9ter1a across authentic tool-calling conversations rendered through the model's native chat template and prose corpus.

EXPLORE THE LFM2.5-24B-A2B EDITIONS

Choose the quantization profile that matches your available memory:

LFM2.5-24B-A2B · APEX-I-MiniPlus V2.1 — 9.64 GB. The balanced performance edition with Q3_K edge experts and Q5_K lead-in armor.

Collections: APEX-I-MiniPlus V2.1 Collection · APEX-I-NanoPlus Collection

Quick Navigation Index

Quantization Metrics & Quality Comparison
Official Liquid AI Benchmarks
Empirically Audited Tensor Precision Map
Model Files & Specifications
Recommended Generation Parameters
Serving & Deployment Setup
Community Compute Fund & Student Support

1. Quantization Metrics & Quality Comparison

Size & Quality Metrics

Edition / Quantization	File Size	Memory Footprint (Weights)	Total Audited Tensors
BF16 Base (Liquid AI)	47.96 GB	44.67 GiB	428
APEX-I-MiniPlus V2.1	9.64 GB	8.98 GiB	428
APEX-I-NanoPlus (This Release)	8.15 GB	7.59 GiB	428

Independent Hardware & Compute Transparency: Generating multi-million token calibration matrices and running certified quantizations on 24B architectures requires dedicated high-end cloud nodes funded 100% out-of-pocket. If you rely on these MiniPlus/NanoPlus builds for your local workloads, contributions to the IsValorum Compute Fund on Ko-fi directly sponsor the GPU/CPU hours required for future Day-0 handcrafted quantizations.

2. Official Liquid AI Benchmarks

LFM2.5-24B-A2B is a 24B parameter reasoning model with only 2B active parameters per token, outperforming larger dense and sparse alternatives across agentic, coding, and mathematical evaluations while running up to 1.7x faster on edge devices:

Key Empirical Strengths:

Autonomous Tool-Use & Function Calling: Reaches state-of-the-art results on LiveClaw (48.09 vs 38.19 on Qwen3.6-35B-A3B) and BFCL v4.
Constant Latency Linear ConvNet Backbone: 30 of the 40 layers utilize 1D short convolutions with O(1) state caching, resulting in blistering inference speeds and near-zero TTFT overhead.
Periodic Attention Anchor: 10 full attention layers distributed every 4 blocks retain long-horizon coherence up to 128,000 tokens.

3. Empirically Audited Tensor Precision Map

Empirical binary audit of all 428 tensors stored within LFM2.5-24B-A2B.APEX-I-NanoPlus.gguf:

========================================================================================
AUDITED GGUF TENSOR ARCHITECTURE (TOTAL STORED TENSORS: 428)
========================================================================================
Format      Count    Share %    Target Components & Architectural Allocation
────────────────────────────────────────────────────────────────────────────────────────
F32           207     48.36%    Zero-drift Routers (ffn_gate_inp), Norms, Biases, Conv Kernels
Q4_K           97     22.66%    Dense FFN Lead-In (0-1), Attention q/k/v, Shortconv, Embeddings
IQ3_XXS        42      9.81%    Border Experts (gate/up: Layers 2,3,38,39) & Core Experts (down: Layers 4-37)
IQ2_S          34      7.94%    Core MoE Experts (gate: Layers 4 to 37)
IQ2_XXS        34      7.94%    Core MoE Experts (up: Layers 4 to 37)
Q6_K           10      2.34%    Attention Output Heads (All 10 Full Attention Layers)
Q3_K            4      0.93%    Border MoE Experts (down: Layers 2, 3, 38, 39)
========================================================================================

Detailed Layer Breakdown:

Layer Range	Block Architecture	Component / Tensor Target	Quantization Format	Design Rationale
Global	Vocab & Output	`token_embd.weight` (Tied output)	`Q4_K`	Preserves core lexical and tokenization fidelity
Global	Embedding Norm	`token_embd_norm.weight`	`F32`	Exact normalizer preservation
Layers 0 & 1	Dense FFN Lead-In	`blk.{0,1}.ffn_{gate,down,up}.weight`	`Q4_K`	Uncompressed 4-bit protection on initial representation blocks
Layers 2 to 39	MoE Routing Gates	`blk.{2..39}.ffn_gate_inp.weight`	`F32`	Absolute zero routing drift across all 64 experts
Layers 0 to 39	Layer Normalizers	`blk.*.{attn_norm,ffn_norm}`	`F32`	Unquantized layer normalizers
Layers 2..39	Expert Biases	`blk.*.exp_probs_b.bias`	`F32`	Exact expert logit bias preservation
10 Layers (2,6,10,14,18,22,26,30,34,38)	Quadratic Attention	`blk.*.attn_{q,k,v}.weight`	`Q4_K`	Robust multi-head projection fidelity up to 128K context
10 Layers (2,6,10,14,18,22,26,30,34,38)	Attention Projection	`blk.*.attn_output.weight`	`Q6_K`	Protected output projections for long-horizon attention
30 Layers (Linear Conv)	Shortconv 1D	`blk.*.shortconv.{in,out}_proj.weight`	`Q4_K`	High-speed linear projection across all 30 non-quadratic blocks
30 Layers (Linear Conv)	Conv Kernels	`blk.*.shortconv.conv.weight`	`F32`	Precision retention for 1D convolution weights
Layers 2, 3, 38, 39 (4 Layers)	Border MoE Down	`blk.*.ffn_down_exps.weight`	`Q3_K`	High stability for critical down-projections at boundary layers
Layers 2, 3, 38, 39 (4 Layers)	Border MoE Gate/Up	`blk.*.ffn_{gate,up}_exps.weight`	`IQ3_XXS`	Balanced importance quantization at ingress and egress
Layers 4–37 (34 Layers)	Core MoE Down	`blk.*.ffn_down_exps.weight`	`IQ3_XXS`	Protected projection matrix for intermediate reasoning tokens
Layers 4–37 (34 Layers)	Core MoE Gate	`blk.*.ffn_gate_exps.weight`	`IQ2_S`	Calibrated 2-bit grid quantization for expert gating
Layers 4–37 (34 Layers)	Core MoE Up	`blk.*.ffn_up_exps.weight`	`IQ2_XXS`	Lightweight up-projection for maximum memory throughput

4. Model Files & Specifications

File Name	Size	Whole-File BPW	Format	Quantization Profile
`LFM2.5-24B-A2B.APEX-I-NanoPlus.gguf`	8.15 GB	aprox. 2.72	GGUF	APEX-I-NanoPlus

5. Recommended Generation Parameters

Official recommended generation hyperparameters as specified in generation_config.json by Liquid AI:

Parameter	Recommended Value	Description
Temperature	`0.1`	Low temperature ensures sharp, factual, and reproducible reasoning trajectories.
Top-K	`50`	Filters candidate tokens to the top 50 logit distribution.
Top-P	`0.95`	Standard nucleus threshold for balanced exploration.
Repetition Penalty	`1.05`	Mild penalty preventing looping without distorting coding indentation.
Max Context	`131,072`	Native context capacity (`128K`).

6. Serving & Deployment Setup

llama.cpp Server (OpenAI-Compatible Endpoint)

llama-server   -m LFM2.5-24B-A2B.APEX-I-NanoPlus.gguf   -c 32768   -ngl 99   --port 8080   --temp 0.1   --top-k 50   --repeat-penalty 1.05

ChatML Prompt Template Structure

<|startoftext|><|im_start|>system
You are a helpful assistant trained by Liquid AI.<|im_end|>
<|im_start|>user
Write a high-performance Python script to parse large JSON streams.<|im_end|>
<|im_start|>assistant
<think>

7. Community Compute Fund & Student Support

All IsValorum quantizations will always remain completely free and open to the public without paywalls.

However, cloud GPU and CPU-performance compute is expensive and funded entirely out-of-pocket. If you find these handcrafted builds to be of high quality and they benefit your day-to-day workflow, you can support future Day-0 releases on Ko-fi. Every contribution directly funds the cloud compute hours required for new model runs and helps support my university education:

(When supporting on Ko-fi, feel free to leave a note with your Hugging Face handle and the specific model you would like prioritized).