Instructions to use WamphyreAI/LiquidVamp-R4-1.7B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use WamphyreAI/LiquidVamp-R4-1.7B with Transformers:

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

pipe = pipeline("text-generation", model="WamphyreAI/LiquidVamp-R4-1.7B", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("WamphyreAI/LiquidVamp-R4-1.7B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("WamphyreAI/LiquidVamp-R4-1.7B", trust_remote_code=True, 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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use WamphyreAI/LiquidVamp-R4-1.7B with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf WamphyreAI/LiquidVamp-R4-1.7B:BF16
# Run inference directly in the terminal:
llama cli -hf WamphyreAI/LiquidVamp-R4-1.7B:BF16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf WamphyreAI/LiquidVamp-R4-1.7B:BF16
# Run inference directly in the terminal:
llama cli -hf WamphyreAI/LiquidVamp-R4-1.7B:BF16

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf WamphyreAI/LiquidVamp-R4-1.7B:BF16
# Run inference directly in the terminal:
./build/bin/llama cli -hf WamphyreAI/LiquidVamp-R4-1.7B:BF16

Use Docker

docker model run hf.co/WamphyreAI/LiquidVamp-R4-1.7B:BF16

Llama new
LM Studio
Jan

vLLM

How to use WamphyreAI/LiquidVamp-R4-1.7B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "WamphyreAI/LiquidVamp-R4-1.7B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "WamphyreAI/LiquidVamp-R4-1.7B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/WamphyreAI/LiquidVamp-R4-1.7B:BF16

SGLang

How to use WamphyreAI/LiquidVamp-R4-1.7B 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 "WamphyreAI/LiquidVamp-R4-1.7B" \
    --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": "WamphyreAI/LiquidVamp-R4-1.7B",
		"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 "WamphyreAI/LiquidVamp-R4-1.7B" \
        --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": "WamphyreAI/LiquidVamp-R4-1.7B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use WamphyreAI/LiquidVamp-R4-1.7B with Ollama:
```
ollama run hf.co/WamphyreAI/LiquidVamp-R4-1.7B:BF16
```
Unsloth Desktop

How to use WamphyreAI/LiquidVamp-R4-1.7B with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf WamphyreAI/LiquidVamp-R4-1.7B:BF16

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "WamphyreAI/LiquidVamp-R4-1.7B:BF16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use WamphyreAI/LiquidVamp-R4-1.7B with Docker Model Runner:
```
docker model run hf.co/WamphyreAI/LiquidVamp-R4-1.7B:BF16
```

Lemonade

How to use WamphyreAI/LiquidVamp-R4-1.7B with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull WamphyreAI/LiquidVamp-R4-1.7B:BF16

Run and chat with the model

lemonade run user.LiquidVamp-R4-1.7B-BF16

List all available models

lemonade list

Hermes Agent

How to use WamphyreAI/LiquidVamp-R4-1.7B with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf WamphyreAI/LiquidVamp-R4-1.7B:BF16

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default WamphyreAI/LiquidVamp-R4-1.7B:BF16

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use WamphyreAI/LiquidVamp-R4-1.7B with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf WamphyreAI/LiquidVamp-R4-1.7B:BF16

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "WamphyreAI/LiquidVamp-R4-1.7B:BF16" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

LiquidVamp-R4-1.7B

By Wamphyre AI.

LiquidVamp-R4-1.7B is an experimental, abliterated-by-lineage language model based on OrcaRouter's LFM3-1.7B-Abliterated. It combines direct neuronal-level weight edits from Wamphyre AI's experimental development with supervised post-training supported by GLM-5.3-Flash as the synthetic-data teacher. Its adaptation focuses on coding, reasoning and tool use.

Abliterated base and model lineage

The upstream base is orcarouter/LFM3-1.7B-abliterated, derived from Liquid AI's LFM3 family. LiquidVamp inherits the abliterated base out of the box. Wamphyre AI's preceding experimental revisions include direct edits to FFN weights at the level of neurons and projection columns.

R4 adds supervised LoRA post-training using GLM-5.3-Flash responses. The adapter has been merged into the released weights. The original LFM3 architecture, vocabulary and conversation template are retained.

High-context local inference

The local deployment is designed for high-context use and configured for a 131,072-token total context window (128K). This window is shared by the prompt, conversation history, optional reasoning and generated output; the configured output budget is 16,384 tokens.

The validated local backend is llama.cpp b12179, using Vulkan/RADV, on Ubuntu with an AMD Radeon RX 6600 8 GB. The deployment supports n-gram speculative decoding through ngram-mod, configured with a 24-token match and 8-16-token draft proposals.

Service initialization at 128K and basic API interaction were verified. The scored evaluation used 8,192 tokens of context, so the results below do not establish answer quality or throughput across a full 128K conversation. The technical specification records the exact runtime settings and verification scope.

Post-training

Item	Value
Model family size	Approximately 1.7 billion parameters
Teacher	GLM-5.3-Flash
Unique teacher exercises	2,163
Training / validation / held-out split	1,362 / 413 / 388
Balanced rows per epoch	1,611
Trainable LoRA parameters	9,999,872
Completed optimization	1 epoch, 101 steps
Training precision and hardware	FP32, NVIDIA Tesla T4
Objective	Loss on assistant responses and tool calls only
Mean training loss	0.63194
Final validation loss	0.53692

The final training run had finite gradients and verified adapter updates. Merging updated 142 projections; the other 145 GGUF tensors retained their previous values exactly.

Local evaluation

Both LiquidVamp R4 and its modified predecessor R3 completed the same 436-exercise battery under matched inference settings, one model loaded at a time. R3 is an OrcaRouter-derived experimental revision with prior Wamphyre AI modifications; these are R3 measurements, not measurements of the untouched upstream OrcaRouter checkpoint.

Test	Modified predecessor R3	LiquidVamp R4
Held-out set	318/388 (81.96%)	357/388 (92.01%)
Held-out Python	9/58	44/58
Held-out reasoning	108/129	112/129
Held-out tools, simulated observations	201/201	201/201
Fresh transfer tasks	14/32	13/32
Entire local battery	343/436	382/436

The paired balance was 56 improvements and 17 regressions. Python gains include improved executable-code formatting: held-out failures containing syntax or indentation errors fell from 45 to 1. This does not establish a purely algorithmic improvement.

Fresh transfer did not improve, and aggregate JavaScript results fell from 3/5 to 1/5. Tool observations in the held-out tests were simulated. Related exercise families limit the independence of the samples. This is a local evaluation, not a standard external benchmark, and the model remains experimental.

The untouched OrcaRouter base has not been evaluated on this battery. No scores or percentage gains against it are estimated. A numeric-punctuation scoring issue was corrected and the same corrected checks were reapplied to both models' saved answers without generation retries. The machine-readable record is training_summary.json.

Release formats

Full merged Safetensors checkpoint, including configuration, tokenizer and architecture source files. It supports inference and further training under the LiquidVamp identity.
BF16 GGUF, with inherited F32 tensors preserved, for compatible local inference runtimes.

The LoRA update is already merged. Loading the complete checkpoint requires no additional adapter or original R3 checkpoint. Transformers loading uses the included architecture code through trust_remote_code=True.

SHA256SUMS contains the release file hashes.

Attribution and license

Wamphyre AI: experimental neuronal edits, adaptation and post-training. OrcaRouter: abliterated upstream base. Liquid AI: original LFM3 family. GLM-5.3-Flash: synthetic-data teacher.

The inherited LFM Open License v1.0 is included in LICENSE. Original notices and modification details are retained in NOTICE.txt.

Downloads last month: -

Safetensors

Model size

2B params

Tensor type

BF16

Model tree for WamphyreAI/LiquidVamp-R4-1.7B

Base model

LiquidAI/LFM3-1.7B-Base

Finetuned

LiquidAI/LFM3-1.7B

Finetuned

orcarouter/LFM3-1.7B-abliterated

Finetuned

(2)

this model