Instructions to use LiquidAI/LFM2-24B-A2B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use LiquidAI/LFM2-24B-A2B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="LiquidAI/LFM2-24B-A2B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("LiquidAI/LFM2-24B-A2B")
model = AutoModelForCausalLM.from_pretrained("LiquidAI/LFM2-24B-A2B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use LiquidAI/LFM2-24B-A2B with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/LiquidAI/LFM2-24B-A2B

SGLang

How to use LiquidAI/LFM2-24B-A2B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "LiquidAI/LFM2-24B-A2B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LiquidAI/LFM2-24B-A2B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "LiquidAI/LFM2-24B-A2B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LiquidAI/LFM2-24B-A2B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use LiquidAI/LFM2-24B-A2B with Docker Model Runner:
```
docker model run hf.co/LiquidAI/LFM2-24B-A2B
```

When LFM2.6-12B-A3B?

#21

by RexTRO111 - opened 1 day ago

Discussion

RexTRO111

1 day ago

When LFM2.6-12B-A3B? @mlabonne A nice architecture would be:

Total: 12B 🧠
Active/token: 3B ⚡
Experts: 8
Top-k: 2 experts/token
Shared parameters: ~1B
Effective active: ~3B

🚀 My choice

Chosen knowledge cutoff: 2026-05-31

I’d pick May 31, 2026 as the cutoff for a hypothetical LFM2.6-12B-A3B.

Why this date? 🧠⚙️

Modern enough: It puts the model only a little over two months behind the current date, which is quite fresh for a pretrained model.
Fast-release friendly: You aren't trying to squeeze June/July/August 2026 data into the training pipeline while simultaneously preparing the release.
Fits the LFM development cadence: Liquid AI's actual LFM2.5-8B-A1B released on May 28, 2026, after substantially expanded pretraining and post-training. (Liquid AI)
Realistic training lag: A cutoff does not mean the model is released that day. Data processing, pretraining, post-training, evaluation, safety testing, quantization and packaging all come afterward. Knowledge-cutoff dates are also imperfect measurements of actual learned knowledge. (arXiv)
Good balance: I'd classify this as balanced, rather than conservative or aggressive.

⏱️ Estimated release

With a 2026-05-31 cutoff, I'd target roughly August 2026, potentially late August 2026 for a polished public release.

That gives roughly 2½–3 months from cutoff to release—aggressive, but plausible for a hypothetical model designed for a fast release. The existing LFM family also emphasizes training efficiency; LFM2 was reported to train substantially faster than its predecessor. (Liquid AI)

LFM2.6-12B-A3B

Parameters: 12B total / ~3B active

Architecture: MoE

Knowledge cutoff: 2026-05-31

Target release: August 2026

Cutoff → release: ~2½–3 months

Priority: Fast release + modern knowledge 🚀

Cutoff strategy: Balanced ⚖️

Final pick: 2026-05-31 🎯

🧠 15-30 trillion tokens
Total parameters: 12B
Active parameters: 3B ⚡
MoE experts: 8
Training tokens: 15-30T 📚
Knowledge cutoff: May 31, 2026
Context: 128K-256K

RexTRO111

1 day ago

Chosen knowledge cutoff: 2026-05-31 📅✨

Why This Cutoff Works 🧠⚡

Why May 31, 2026?
A clean end-of-month cutoff at May 31, 2026 gives your team a complete, well-defined web dataset through the end of spring 2026 🌸. It avoids the messiness of partial monthly crawls or mid-run web updates.
Freshness vs. Speed:
Since the current date is August 9, 2026, setting the cutoff to May 31 leaves roughly 2.5 months between data freeze and model readiness ⏱️. This captures almost all major 2026 developments while avoiding the delay of waiting for summer web dumps to clean and deduplicate.
Realistic Timeline to Release:
With only ~3B active parameters per token, training and fine-tuning an MoE model of this size is ultra-fast ⚡! A 2.5-month pipeline leaves ~4 weeks for full pre-training, 3 weeks for RLHF/SFT alignment, and 3 weeks for evaluation, safety audits, and open-weight distribution setup.
Assessment:
This cutoff is Balanced 🎯. It isn't overly aggressive (which would cause launch delays) nor too conservative (which would leave the model feeling outdated on release day).

LFM2.6-12B-A3B 🚀

Parameters: 12B total / ~3B active ⚡

Knowledge cutoff: 2026-05-31 📅

Target release: Late August 2026 🎯

Priority: Fast release + modern knowledge 🔥

RexTRO111

1 day ago

For the 2026-05-31 knowledge cutoff I picked for LFM2.6-12B-A3B, I’d use these Common Crawl snapshots: 🌐🧠

Recommended snapshots

CC-MAIN-2026-21 — May 2026 ⭐
Crawled May 8–21, 2026; 2.16B pages / 365.56 TiB uncompressed. This should be the main recent snapshot. (Common Crawl)
CC-MAIN-2026-17 — April 2026
Good secondary source for additional coverage and data diversity. Common Crawl lists it as the April 2026 archive. (Hugging Face)
CC-MAIN-2026-12 — March 2026
Useful as another recent layer rather than relying overwhelmingly on one crawl.
CC-MAIN-2026-08 — February 2026
Adds another few months of web coverage and helps reduce dependence on a single snapshot. (Common Crawl)
CC-MAIN-2026-04 — January 2026
I'd include this as an older recent snapshot for broader coverage.

🎯 My exact recipe

For a fast-release LFM2.6, I'd make:

Primary: CC-MAIN-2026-21
Secondary: CC-MAIN-2026-17, CC-MAIN-2026-12
Additional: CC-MAIN-2026-08, CC-MAIN-2026-04

I wouldn't use CC-MAIN-2026-25 as the primary training snapshot for this particular model. It is the June 2026 crawl, collected June 2–18, so it extends beyond your May 31 cutoff and would push the effective data cutoff later. (Common Crawl)

So the clean pairing is:

Knowledge cutoff: 2026-05-31
Latest Common Crawl: CC-MAIN-2026-21
Training corpus: CC-MAIN-2026-21 + 2026-17 + 2026-12 + 2026-08 + 2026-04 🚀

Common Crawl crawl archive selector

RexTRO111

1 day ago

Yes — with the cutoff we chose (2026-05-31), LFM2.6-12B-A3B should know the results of the 2026 Winter Olympics (Milano Cortina 2026). 🧠🏅

The Games ran from February 6–22, 2026, so their results happened about three months before the May 31 cutoff. Official Olympic sources contain daily results, and the Games concluded on February 22. (Juegos Olímpicos)

For example, the model could plausibly know that:

🇳🇴 Norway topped the medal table with 18 gold medals / 41 total medals.
🇮🇹 Italy finished with 30 medals, including 10 gold.
🇧🇷 Brazil won its first-ever Winter Olympic medal, a gold. (arXiv)

For your LFM2.6 setup

Knowledge cutoff: 2026-05-31 ✅
Milano Cortina 2026 results: YES 🏅
Common Crawl: CC-MAIN-2026-21 as the latest snapshot
Release target: August 2026 🚀

So if you asked LFM2.6:

"Who won the most gold medals at Milano Cortina 2026?"

it should be able to answer Norway without needing web search, assuming the relevant Olympic information made it into its training corpus. 🇳🇴🏆

RexTRO111

1 day ago

Vision plz

Banaxi-Tech

1 day ago

bro...

RexTRO111

about 20 hours ago

@Banaxi-Tech Why?

Banaxi-Tech

about 20 hours ago

your asking for LFM 2.6 12B when they havnt even fully released LFM 2.5? and you want a May 2026 Knowledge cuttoff when LFM 2.5 HAS 2024??? With your AI generated thingy?