batiai/ lfm2.5-8b

54 Downloads Updated 1 week ago

ollama run batiai/lfm2.5-8b:iq3

curl http://localhost:11434/api/chat \
  -d '{
    "model": "batiai/lfm2.5-8b:iq3",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='batiai/lfm2.5-8b:iq3',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'batiai/lfm2.5-8b:iq3',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Models

Name

6 models

Size / Usage

Context

Input

lfm2.5-8b:iq3

3.3GB · 125K context window · Text · 1 week ago

lfm2.5-8b:iq3

3.3GB

125K

Text

lfm2.5-8b:iq4

4.6GB · 125K context window · Text · 1 week ago

lfm2.5-8b:iq4

4.6GB

125K

Text

lfm2.5-8b:q2

3.0GB · 125K context window · Text · 1 week ago

lfm2.5-8b:q2

3.0GB

125K

Text

lfm2.5-8b:q3

4.1GB · 125K context window · Text · 1 week ago

lfm2.5-8b:q3

4.1GB

125K

Text

lfm2.5-8b:q4

5.2GB · 125K context window · Text · 1 week ago

lfm2.5-8b:q4

5.2GB

125K

Text

lfm2.5-8b:q6

7.0GB · 125K context window · Text · 1 week ago

lfm2.5-8b:q6

7.0GB

125K

Text

Readme

Liquid LFM2.5-8B-A1B — Quantized by BatiAI

Liquid AI’s on-device MoE — 8.3B total / 1.5B active (32 experts, 4 active). Reasoning + native tool calling, 128K context. Built for the edge: hybrid LIV-conv + GQA architecture, 253 tok/s on M5 Max under 6GB. Quantized directly from official Liquid weights.

Models

Tag	Size	RAM target	Use Case
q2	~2.8GB	8GB Mac	Ultra-compact (imatrix)
iq3	~3.2GB	8GB Mac	imatrix, smallest K-class
q3	~3.9GB	8GB+ Mac	Balanced
iq4	~4.3GB	16GB Mac	imatrix, best size/quality
q4	~4.9GB	16GB Mac	Recommended
q6	~6.5GB	16GB+	Near-original

Quick Start

ollama run batiai/lfm2.5-8b:q4

Why LFM2.5-8B-A1B?

1.5B active params — 8B-class quality at ~1.5B-class speed
Hybrid edge architecture — 18 double-gated LIV conv + 6 GQA layers (not a GPU-first transformer)
Reasoning — explicit chain-of-thought before final answer
Native tool calling — Pythonic or JSON function calls
128K context, 38T-token training, 128K vocab (better Korean/CJK tokenization)
Released May 28, 2026

RAM Requirements

Your Mac RAM	iq3	q2	q3	iq4	q4	q6
8GB	✅	✅	✅	✅	⚠️	❌
16GB	✅	✅	✅	✅	✅	✅
24GB+	✅	✅	✅	✅	✅	✅

For Other Macs

Your Mac	Recommended
8GB	`batiai/lfm2.5-8b:q3` (this) or `batiai/gemma4-e2b:q4`
16GB	`batiai/lfm2.5-8b:q4` (this, recommended) or `batiai/granite4.1:q4`
24GB	`batiai/gemma4-26b:iq4`
32GB	`batiai/nemotron3-nano:iq4`
96GB+	`batiai/mistral-medium-3.5:iq3`

Why BatiAI?

Quantized directly from official Liquid weights
imatrix calibrated (IQ variants)
Low quants (q2/q3/iq3) the official repo doesn’t ship — for 8GB Macs
Tool calling validated
BatiAI signed (general.author=BatiAI)

License

LFM Open License v1.0 — https://www.liquid.ai/lfm-license. Free commercial use under $10M USD annual revenue. Original Liquid AI attribution retained.

Built for BatiFlow

Free, on-device AI automation for Mac. 5MB app, 100% local, unlimited.

https://flow.bati.ai

# Liquid LFM2.5-8B-A1B — Quantized by BatiAI

Liquid AI's on-device MoE — **8.3B total / 1.5B active** (32 experts, 4 active). Reasoning + native tool calling, 128K context.
Built for the edge: hybrid LIV-conv + GQA architecture, 253 tok/s on M5 Max under 6GB.
Quantized directly from [official Liquid weights](https://huggingface.co/LiquidAI/LFM2.5-8B-A1B).

## Models

| Tag | Size | RAM target | Use Case |
|-----|------|------------|----------|
| **q2** | ~2.8GB | 8GB Mac | Ultra-compact (imatrix) |
| **iq3** | ~3.2GB | 8GB Mac | imatrix, smallest K-class |
| **q3** | ~3.9GB | 8GB+ Mac | Balanced |
| **iq4** | ~4.3GB | 16GB Mac | imatrix, best size/quality |
| **q4** | ~4.9GB | 16GB Mac | **Recommended** |
| **q6** | ~6.5GB | 16GB+ | Near-original |

## Quick Start

```
ollama run batiai/lfm2.5-8b:q4
```

## Why LFM2.5-8B-A1B?

- **1.5B active params** — 8B-class quality at ~1.5B-class speed
- **Hybrid edge architecture** — 18 double-gated LIV conv + 6 GQA layers (not a GPU-first transformer)
- **Reasoning** — explicit chain-of-thought before final answer
- **Native tool calling** — Pythonic or JSON function calls
- **128K context**, 38T-token training, 128K vocab (better Korean/CJK tokenization)
- Released May 28, 2026

## RAM Requirements

| Your Mac RAM | iq3 | q2 | q3 | iq4 | q4 | q6 |
|-------------|-----|-----|-----|-----|-----|-----|
| **8GB** | ✅ | ✅ | ✅ | ✅ | ⚠️ | ❌ |
| **16GB** | ✅ | ✅ | ✅ | ✅ | **✅** | ✅ |
| 24GB+ | ✅ | ✅ | ✅ | ✅ | ✅ | **✅** |

## For Other Macs

| Your Mac | Recommended |
|----------|-------------|
| 8GB | `batiai/lfm2.5-8b:q3` (this) or `batiai/gemma4-e2b:q4` |
| 16GB | `batiai/lfm2.5-8b:q4` (this, recommended) or `batiai/granite4.1:q4` |
| 24GB | `batiai/gemma4-26b:iq4` |
| 32GB | `batiai/nemotron3-nano:iq4` |
| 96GB+ | `batiai/mistral-medium-3.5:iq3` |

## Why BatiAI?

- Quantized directly from official Liquid weights
- imatrix calibrated (IQ variants)
- Low quants (q2/q3/iq3) the official repo doesn't ship — for 8GB Macs
- Tool calling validated
- BatiAI signed (`general.author=BatiAI`)

## License

LFM Open License v1.0 — https://www.liquid.ai/lfm-license. Free commercial use under $10M USD annual revenue. Original Liquid AI attribution retained.

## Built for BatiFlow

Free, on-device AI automation for Mac. 5MB app, 100% local, unlimited.

https://flow.bati.ai

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