Instructions to use cstr/Phoenix-laser with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use cstr/Phoenix-laser with Transformers:

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

pipe = pipeline("text-generation", model="cstr/Phoenix-laser")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("cstr/Phoenix-laser")
model = AutoModelForCausalLM.from_pretrained("cstr/Phoenix-laser", 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 cstr/Phoenix-laser with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/cstr/Phoenix-laser

SGLang

How to use cstr/Phoenix-laser 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 "cstr/Phoenix-laser" \
    --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": "cstr/Phoenix-laser",
		"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 "cstr/Phoenix-laser" \
        --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": "cstr/Phoenix-laser",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use cstr/Phoenix-laser with Docker Model Runner:
```
docker model run hf.co/cstr/Phoenix-laser
```

Phoenix-laser / README.md

cstr

docs: add model card — base model, licence, Art. 53 provenance

1c9f30b verified about 1 month ago

preview code

Raw

History Blame Contribute Delete

3.29 kB

	---
	license: apache-2.0
	base_model: DRXD1000/Phoenix-7B
	base_model_relation: finetune
	library_name: transformers
	pipeline_tag: text-generation
	tags:
	- laserRMT
	- mistral
	---

	# Phoenix-laser

	A LaserRMT modification of [`DRXD1000/Phoenix-7B`](https://huggingface.co/DRXD1000/Phoenix-7B).

	LaserRMT (layer-selective rank reduction) alters the base model's existing
	weight matrices. It is not a format conversion and not a quantisation: the
	weights differ from the base in substance, not merely in numeric
	representation.

	Note: this repository the base was published as `DRXD1000/Phoenix` and has since been renamed `DRXD1000/Phoenix-7B`.

	\| \| \|
	\|---\|---\|
	\| Base model \| [`DRXD1000/Phoenix-7B`](https://huggingface.co/DRXD1000/Phoenix-7B) \|
	\| Licence \| `apache-2.0`, inherited from the base \|
	\| Architecture \| MistralForCausalLM, 32 layers (unchanged from the base) \|
	\| Modification \| LaserRMT layer-selective rank reduction \|

	## Provenance and EU AI Act Art. 53 note

	This card was written on 2026-08-02. The repository had carried **no model card
	at all** since it was created on 2024-03-18 — modified weights published with no
	attribution to the model they were derived from. That is the gap this card
	closes, and it is worth stating plainly rather than quietly backfilling.

	The base model above was not guessed from the repository name. It is recorded
	in this repo's own `config.json` as `_name_or_path`, and the licence is the one
	the base declares on the Hub as of 2026-08-02.

	Provider status. Most `cstr/*` repositories are format conversions, where
	the upstream research team remains the provider of the model under Regulation
	(EU) 2024/1689 and the conversion changes only the numeric representation. **This
	repository is not one of those.** Modifying weights places a new model on the
	market, so the obligations that survive the Art. 53(2) free-and-open-source
	exemption — Art. 53(1)(c) and 53(1)(d) — attach here.

	Art. 53(1)(c) — copyright policy. This repository does not introduce any
	training corpus of its own, so no text or data mining was carried out here and
	no rights reservations under Art. 4(3) of Directive (EU) 2019/790 were engaged
	by this step. The modification operates on weights already published by the base
	model's authors under `apache-2.0`. Where the base model's own training raises
	copyright questions, those attach to the base model's provider, whose
	documentation is linked above. Any credible claim that this repository
	redistributes material it has no right to redistribute will be acted on —
	contact via the Community tab.

	Art. 53(1)(d) — training content. No additional training corpus was
	introduced by this repository. The model's training content is that of
	[`DRXD1000/Phoenix-7B`](https://huggingface.co/DRXD1000/Phoenix-7B), and its documentation is the summary
	that applies. This repository does not restate it, because restating training
	data documentation second-hand is how inaccurate provenance claims propagate.

	Known limitation. The precise LaserRMT configuration used in 2024 — which
	layers were reduced, and to what rank — is not recorded in this repository. It
	is not reconstructible from the weights alone with confidence, and it is not
	asserted here rather than being guessed at.