Instructions to use LiquidAI/LFM2.5-2.6B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use LiquidAI/LFM2.5-2.6B with Transformers:

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

pipe = pipeline("text-generation", model="LiquidAI/LFM2.5-2.6B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("LiquidAI/LFM2.5-2.6B")
model = AutoModelForCausalLM.from_pretrained("LiquidAI/LFM2.5-2.6B", 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/LFM2.5-2.6B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "LiquidAI/LFM2.5-2.6B"
# 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/LFM2.5-2.6B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/LiquidAI/LFM2.5-2.6B

SGLang

How to use LiquidAI/LFM2.5-2.6B 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/LFM2.5-2.6B" \
    --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/LFM2.5-2.6B",
		"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/LFM2.5-2.6B" \
        --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/LFM2.5-2.6B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use LiquidAI/LFM2.5-2.6B with Docker Model Runner:
```
docker model run hf.co/LiquidAI/LFM2.5-2.6B
```

Request for Historical Evaluation Manifests for LFM2.5 Benchmark Reproduction

#15

by chase-jay - opened 10 days ago

Discussion

chase-jay

10 days ago

Hello Liquid AI team,

We are independently reproducing the benchmark results published for LFM2.5-2.6B and the accompanying Gemma-4 and Qwen3.5 comparisons.

We treat the published values only as comparison targets. Our reproduced scores are recorded separately, and every undisclosed field remains marked as UNKNOWN rather than being inferred as an official setting.

Could you please share or confirm the exact historical evaluation manifests used for the published tables?

Global runtime
Exact model revisions
vLLM, PyTorch, and CUDA versions/commits
GPU type, dtype, tensor parallelism, and quantization
gpu_memory_utilization, concurrency, and scheduling configuration
Complete model-server commands
Generation configuration
Full prompt/chat template for each benchmark
Thinking/reasoning settings
Reasoning and tool parsers
temperature, top_p, top_k, min_p, and penalties
Seed, number of runs, retries, stop sequences, and timeout
Benchmark-specific manifests
BFCLv4: effective temperature, runtime, handlers for all five models, repository commit, aggregation procedure, and raw category CSVs
IFStruct: evaluator/dataset commits, Qwen thinking setting, reasoning parser, and whether the validator received raw reasoning or final-answer text
LiveCodeBench v6: prompt style, top_p, seed, and exact repository revision
ToolSandbox: scenario list, user simulator, search backend, tool handler, and aggregation procedure
BrowseComp+ and Claw-Eval: exact OpenClaw version/configuration, task list, retrieval/index settings, graders, judges, and trial aggregation
Multi-IF, PinchBench, τ³-Bench, and AA Omniscience: exact historical revisions and run manifests

Two ambiguities we observed appear to be particularly score-critical.

First, Liquid's public BFCL stack requests temperature=0.001, while its pinned vLLM version clamps positive temperatures below 0.01 to 0.01. Our complete public-stack run contains 5,217/5,217 valid rows and scores 49.43%, compared with the published 56.88%.

Second, using the same Qwen3.5-4B checkpoint and seeds, we observed IFStruct scores of 86.00%, 61.50%, and 16.00% solely from changes to the thinking/parser response-boundary settings. The published value is 36.25%, so the exact historical generation and evaluation boundary is essential for reproducibility.

Redacted manifests, server commands, handler definitions, raw generations, or evaluator summaries would all be sufficient. If any details cannot be disclosed, confirmation of that would also be useful; we will keep those fields marked as UNKNOWN.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment