---
license: other
license_link: LICENSE
thumbnail: https://huggingface.co/AtomicChat/lfm25-8b-a1b-GGUF/resolve/main/hero.png
base_model:
- LiquidAI/LFM2.5-8B-A1B
base_model_relation: quantized
quantized_by: AtomicChat
pipeline_tag: text-generation
library_name: gguf
tags:
- atomic-chat
- lfm2.5
- liquidai
- gguf
- llama.cpp
- quantized
---
**LFM2.5 8B A1B**, self-quantized to GGUF by [Atomic Chat](https://atomic.chat). Built straight from Liquid AI's original weights with a per-tensor importance matrix, so this is not a repack of somebody else's files. Runs fully offline.
## Highlights
- **8.5B parameters**: the weights this repo quantizes.
- **Context length**: 128,000 tokens (125K), as published by Liquid AI.
- **24 layers**: Mixture-of-Experts.
- **Full imatrix ladder**: every quant is calibrated with an importance matrix.
- **On-device personal assistant**: Designed to power real-life applications, chaining tool calls, and following complex instructions on all devices.
- **Compressed performance**: Competitive with much larger dense and MoE models on instruction following and agentic tasks.
- **Unmatched throughput**: Fastest in its size class on both CPU and GPU inference, with day-one support for llama.cpp, MLX, vLLM, and SGLang.
> [!NOTE]
> These GGUFs are **self-quantized from the original weights**, not a repack. The importance matrix keeps low-bit quants closer to the full-precision model.
> [!IMPORTANT]
> Always pass `--jinja` so the **LFM2.5 8B A1B chat template** is applied. Without it the model can emit malformed turns.
## Model Overview
| Property | Value |
|---|---|
| Base model | `LiquidAI/LFM2.5-8B-A1B` |
| Parameters | 8.5B |
| Layers | 24 |
| Experts | 32 routed (top-4) |
| Context length | 128,000 tokens (125K) |
| Vocabulary | 128,000 |
| Modalities | Text |
| Architecture | Mixture-of-Experts, 32 experts (top-4), 32 attention heads over 8 KV heads, `Lfm2MoeForCausalLM` |
| This repo | GGUF quants (imatrix). Quants: `Q2_K`, `IQ3_M`, `Q3_K_M`, `Q3_K_L`, `IQ4_XS`, `Q4_K_S`, `Q4_K_M`, `UD-Q4_K_XL`, `Q5_K_S`, `Q5_K_M`, `Q6_K`, `Q8_0` |
Scores are Liquid AI's published results for the base `LiquidAI/LFM2.5-8B-A1B`, not our own measurements. Quantization preserves the large majority of this; `Q4_K_M` and up stay close to full precision.
## Choosing a quant
| Quant | Size | Notes |
|---|---|---|
| `Q2_K` | 3.2 GB | Smallest K-quant. Minimal RAM, clear quality drop. |
| `IQ3_M` | 3.8 GB | Beats Q3 at a similar size thanks to imatrix. Best low-RAM pick. |
| `Q3_K_M` | 4.1 GB | Low quality but usable. |
| `Q3_K_L` | 4.4 GB | A step above Q3_K_M. |
| `IQ4_XS` | 4.6 GB | Excellent quality for size. Recommended low-bit. |
| `Q4_K_S` | 4.9 GB | Compact 4-bit, fast. |
| **`Q4_K_M`** | 5.2 GB | **Recommended default. Best balance of size, speed and quality.** |
| `UD-Q4_K_XL` | 5.2 GB | Dynamic. Embeddings and output kept at Q8_0 for higher quality at a Q4 footprint. |
| `Q5_K_S` | 5.9 GB | Higher quality, slightly more compact than Q5_K_M. |
| `Q5_K_M` | 6.0 GB | Higher quality, low loss. |
| `Q6_K` | 7.0 GB | Near lossless, noticeably lighter than Q8_0. |
| `Q8_0` | 9.0 GB | Effectively lossless, reference quality. |
> [!TIP]
> Pick the largest file that fits your (V)RAM with room for context. `Q4_K_M` or `UD-Q4_K_XL` is the sweet spot for most setups; `Q6_K` or `Q8_0` for maximum fidelity.
## Get started
Run LFM2.5 8B A1B locally with:
- **[Atomic Chat](https://atomic.chat):** the easiest path. Open the app, search `AtomicChat/lfm25-8b-a1b-GGUF`, pick a quant, hit **Use this model**.
- **llama.cpp:** `llama-server -hf AtomicChat/lfm25-8b-a1b-GGUF:Q4_K_M --jinja -c 8192`
- **Ollama:** `ollama run hf.co/AtomicChat/lfm25-8b-a1b-GGUF:Q4_K_M`
- **LM Studio / Jan:** search the repo id, download any quant.
## Best practices
| Parameter | Value |
|---|---|
| temperature | 0.2 |
| top_k | 80 |
| repetition_penalty | 1.05 |
Liquid AI's recommended sampling configuration for `LiquidAI/LFM2.5-8B-A1B`.
## Run in llama.cpp
```bash
git clone https://github.com/ggml-org/llama.cpp
cmake llama.cpp -B llama.cpp/build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON
cmake --build llama.cpp/build --config Release -j --target llama-cli llama-server
```
```bash
./llama.cpp/build/bin/llama-server \
-hf AtomicChat/lfm25-8b-a1b-GGUF:Q4_K_M \
--jinja -ngl 99 -c 8192 -fa on
```
## How these were made
1. Download `LiquidAI/LFM2.5-8B-A1B` (original weights).
2. Convert to f16 GGUF with [llama.cpp](https://github.com/ggml-org/llama.cpp).
3. Build an importance matrix over our calibration corpus.
4. Quantize the ladder with `--imatrix`.
5. `UD-Q4_K_XL` additionally pins the token-embedding and output tensors to `Q8_0`.
## License
Original model by Liquid AI, released under the other license. Full terms: [other](LICENSE). Quantized by Atomic Chat.
