Instructions to use XtrCom/xt-lunaris-checkpoints with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use XtrCom/xt-lunaris-checkpoints with Transformers:

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

pipe = pipeline("text-generation", model="XtrCom/xt-lunaris-checkpoints")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("XtrCom/xt-lunaris-checkpoints", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use XtrCom/xt-lunaris-checkpoints 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 XtrCom/xt-lunaris-checkpoints:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf XtrCom/xt-lunaris-checkpoints:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf XtrCom/xt-lunaris-checkpoints:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf XtrCom/xt-lunaris-checkpoints:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf XtrCom/xt-lunaris-checkpoints:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf XtrCom/xt-lunaris-checkpoints:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf XtrCom/xt-lunaris-checkpoints:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf XtrCom/xt-lunaris-checkpoints:Q4_K_M

Use Docker

docker model run hf.co/XtrCom/xt-lunaris-checkpoints:Q4_K_M

LM Studio
Jan

vLLM

How to use XtrCom/xt-lunaris-checkpoints with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/XtrCom/xt-lunaris-checkpoints:Q4_K_M

SGLang

How to use XtrCom/xt-lunaris-checkpoints 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 "XtrCom/xt-lunaris-checkpoints" \
    --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": "XtrCom/xt-lunaris-checkpoints",
		"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 "XtrCom/xt-lunaris-checkpoints" \
        --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": "XtrCom/xt-lunaris-checkpoints",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use XtrCom/xt-lunaris-checkpoints with Ollama:
```
ollama run hf.co/XtrCom/xt-lunaris-checkpoints:Q4_K_M
```
Unsloth Desktop

How to use XtrCom/xt-lunaris-checkpoints with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf XtrCom/xt-lunaris-checkpoints:Q4_K_M

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": "XtrCom/xt-lunaris-checkpoints:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use XtrCom/xt-lunaris-checkpoints with Docker Model Runner:
```
docker model run hf.co/XtrCom/xt-lunaris-checkpoints:Q4_K_M
```

Lemonade

How to use XtrCom/xt-lunaris-checkpoints with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull XtrCom/xt-lunaris-checkpoints:Q4_K_M

Run and chat with the model

lemonade run user.xt-lunaris-checkpoints-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use XtrCom/xt-lunaris-checkpoints with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf XtrCom/xt-lunaris-checkpoints:Q4_K_M

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 XtrCom/xt-lunaris-checkpoints:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use XtrCom/xt-lunaris-checkpoints with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf XtrCom/xt-lunaris-checkpoints:Q4_K_M

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 "XtrCom/xt-lunaris-checkpoints:Q4_K_M" \
  --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"

XT-Lunaris v2 — abliterated MiniCPM5-2B + LoRA + GRPO

A 2.52B on-device assistant made by XT-Research. Chain: abliteration (a1.0_all) → SFT LoRA → GRPO stages 1–3 → full-slice replay (revision pass). Every adapter stage is preserved at XtrCom/xt-lunaris-checkpoints.

Training progress

run	steps	tokens / rollouts	loss / reward	time
SFT LoRA (r=32)	460	5.66M tokens	9.2 → ~1.7	~30 min
GRPO stage 1 (think tiers, k=4)	20	80 rollouts	R +2.35..+2.98	6.7 min
GRPO stage 2 (continuation)	16	64 rollouts	R +0.75..+2.24	5.4 min
GRPO stage 3 (+ compliance tiers)	20	80 rollouts	R +0.75..+2.98	8.8 min
Replay / revision (full slice, lr 5e-5)	140	~0.57M tokens	1.28 → 0.88–0.91	6.7 min
v3 topup (hellaswag-train + code + compliance + anchors)	120	~6,060 rows	0.57–0.90	5.9 min

Loops across 224 RL rollouts: 0. Think length scales by tier (easy ~62–90, medium ~129–146, hard ~286–363 chars, greedy) — measured, not capped. The v3 topup targets the three measured weak spots: HellaSwag, coding depth, and the refusal markers that survived the slice filters.

Loops across 224 RL rollouts: 0. Think length scales by tier (easy ~62–90, medium ~129–146, hard ~286–363 chars, greedy) — measured, not capped.

Benchmarks — measured this session

Capability battery (repo harness, StreamingLlama engine, 240 letter-scored MC items)

metric	base SFT	abliterated (a1.0_all)	XT-Lunaris v2
arc	0.375	0.4625	0.7375
mmlu	0.4125	0.4625	0.6000
hellaswag	0.300	0.275	0.3250
cap_score	0.3625	0.4000	0.5542
wikitext PPL	11.82	12.22	31.28 (verify — see note)
refusal propensity	0.7010	0.0973	0.4682

PPL note: the code-PPL source returned NaN on this box (dataset absent), so the PPL pipeline is not identical to the session that produced 11.8/12.2 — verify like-for-like before quoting the v2 PPL.

lm-eval 0.4.13, limit 200 per task, greedy, on the replayed chain:

task	metric	score
gsm8k	exact_match, strict-match	0.670
gsm8k	exact_match, flexible-extract	0.445
sciq	acc	0.960
arc_challenge	acc / acc_norm	0.485 / 0.465
hellaswag	acc / acc_norm	0.465 / 0.535

code_eval (generated code actually executed — python asserts in a subprocess, C compiled with gcc): pass@1 = 0.95 (19/20; the single failure is the C maximum program in thinking mode).

v3 topup re-score

PENDING — battery + refusal probe on the v3 chain; lands as the final commit.

Benchmarks — not measured

Terminal-Bench 2.x / 4.0: requires Dockerized task containers and an agent harness. Not runnable in this session's compute window. Do not treat any Terminal-Bench number here as ours.
MMLU via lm-eval: covered by the capability battery's subset instead.

Comparisons — reported by others

Not fetched this session (search integration unavailable; Llama/Gemma cards gated). Fill from the Qwen3 blog before publishing claims about other models.

Quantisations

PENDING — built from lunaris-v2-f16.gguf: Q4_K_M, Q5_K_M (sizes to come).

Known limitations

2.5B scale: loses general-knowledge comparisons to 9B models by arithmetic.
Identity: name binds (XT-Lunaris); maker phrasing still inconsistent.
English-centric; research artifact, not a product.

Downloads last month: -

GGUF

Model size

3B params

Architecture

llama

Hardware compatibility

4-bit

5-bit

16-bit

Model tree for XtrCom/xt-lunaris-checkpoints

Base model

openbmb/MiniCPM5-2B-SFT

Finetuned

XtrCom/MINICPM5-2B-SFT-Abiliterated

Adapter

(1)

this model