sam860/LFM2:2.6b-exp-Q8_0

sam860/ LFM2:2.6b-exp-Q8_0

4,544 Downloads Updated 1 month ago

Small models made to run on mobile devices, developed by Liquid AI

350m 700m 1.2b 2.6b 8b

ollama run sam860/LFM2:2.6b-exp-Q8_0

curl http://localhost:11434/api/chat \
  -d '{
    "model": "sam860/LFM2:2.6b-exp-Q8_0",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='sam860/LFM2:2.6b-exp-Q8_0',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'sam860/LFM2:2.6b-exp-Q8_0',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 1 month ago

1 month ago

a1d8051fa86f · 2.7GB ·

model

archlfm2

·

parameters2.57B

·

quantizationQ8_0

2.7GB

template

{{ if .System }}<|startoftext|><|im_start|>system {{ .System }}<|im_end|> {{ end }}{{ if .Prompt }}<

197B

params

{ "stop": [ "<|startoftext|>", "<|im_start|>", "<|im_end|>", "<|

75B

Readme

Notes

Uploaded as Unsloth DynamicQuant2 (DQ2) versions. These are Liquid AI’s on-device-first models using a hybrid Liquid architecture (gated convolutions + GQA) that’s 2x faster on CPU than standard transformers.

All models here are Q4_0. The unsloth dq2 method may preserve most of the performance of the q8 model, but doesn’t guarantee it. Do your own testing to make sure it really works for your use case.

Temperature: A range of 0.4–0.6 works well for their instruction-following and reasoning tasks.

Description

Liquid Foundation Models 2 (LFM2) — a family of fast, efficient models optimized for edge devices (phones, laptops, vehicles). The hybrid Liquid architecture provides superior speed and memory efficiency over traditional transformers.

The key difference between these models is their architecture and capacity:

Dense Models:

lfm2:350m: Entry-level dense model for ultra-low resource environments.
lfm2:700m: Compact dense model, outperforms Gemma 3 1B IT.
lfm2:1.2b: Dense model that competes with Qwen3-1.7B (47% larger).
lfm2:2.6b: Highest-quality dense model, outperforms 3B+ class models.
lfm2:2.6b-exp: Improved fine-tune of 2.6B for instruction-following, knowledge, and math.

Mixture-of-Experts (On-device MoE):

lfm2:8b-a1b: 8.3B total params but only 1.5B active per token — delivers 8B quality at 1.5B speed with sparse activation using 32 experts (top-4 activated).

Ideal for:

Fast on-device inference (CPU/NPU optimized)
Mobile AI applications and robotics
Low-latency chat and reasoning
English/Japanese-focused tasks with strong multilingual support

References

LFM2 Official Announcement

LFM2-2.6B Details

LFM2-8B-A1B MoE Details

HuggingFace LFM2 Collection

### Notes
Uploaded as **Unsloth DynamicQuant2 (DQ2)** versions. These are Liquid AI's on-device-first models using a hybrid Liquid architecture (gated convolutions + GQA) that's 2x faster on CPU than standard transformers.

All models here are **Q4_0**. The unsloth dq2 method may preserve most of the performance of the q8 model, but doesn't guarantee it. Do your own testing to make sure it really works for your use case.

**Temperature:** A range of **0.4–0.6** works well for their instruction-following and reasoning tasks.

---

### Description
**Liquid Foundation Models 2 (LFM2)** — a family of fast, efficient models optimized for edge devices (phones, laptops, vehicles). The hybrid Liquid architecture provides superior speed and memory efficiency over traditional transformers.

The key difference between these models is their architecture and capacity:

**Dense Models:**

- **lfm2:350m:** Entry-level dense model for ultra-low resource environments.

- **lfm2:700m:** Compact dense model, outperforms Gemma 3 1B IT.

- **lfm2:1.2b:** Dense model that competes with Qwen3-1.7B (47% larger).

- **lfm2:2.6b:** Highest-quality dense model, outperforms 3B+ class models.

- **lfm2:2.6b-exp:** Improved fine-tune of 2.6B for instruction-following, knowledge, and math.

**Mixture-of-Experts (On-device MoE):**

- **lfm2:8b-a1b:** 8.3B total params but only 1.5B active per token — delivers 8B quality at 1.5B speed with sparse activation using 32 experts (top-4 activated).

Ideal for:

- Fast on-device inference (CPU/NPU optimized)

- Mobile AI applications and robotics

- Low-latency chat and reasoning

- English/Japanese-focused tasks with strong multilingual support

---

### References
[LFM2 Official Announcement](https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models)

[LFM2-2.6B Details](https://www.liquid.ai/blog/introducing-lfm2-2-6b-redefining-efficiency-in-language-models)

[LFM2-8B-A1B MoE Details](https://www.liquid.ai/blog/lfm2-8b-a1b-an-efficient-on-device-mixture-of-experts)

[HuggingFace LFM2 Collection](https://huggingface.co/collections/Liquid-AI/lfm-2-series-671a5e7c8e6288d8b6b2c8f5)

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)