Instructions to use texdata/sraka-27b-experimental with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use texdata/sraka-27b-experimental with Transformers:

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

pipe = pipeline("text-generation", model="texdata/sraka-27b-experimental")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("texdata/sraka-27b-experimental")
model = AutoModelForMultimodalLM.from_pretrained("texdata/sraka-27b-experimental", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.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(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use texdata/sraka-27b-experimental 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 texdata/sraka-27b-experimental:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf texdata/sraka-27b-experimental:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf texdata/sraka-27b-experimental:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf texdata/sraka-27b-experimental: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 texdata/sraka-27b-experimental:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf texdata/sraka-27b-experimental: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 texdata/sraka-27b-experimental:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf texdata/sraka-27b-experimental:Q4_K_M

Use Docker

docker model run hf.co/texdata/sraka-27b-experimental:Q4_K_M

LM Studio
Jan

vLLM

How to use texdata/sraka-27b-experimental with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "texdata/sraka-27b-experimental"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "texdata/sraka-27b-experimental",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/texdata/sraka-27b-experimental:Q4_K_M

SGLang

How to use texdata/sraka-27b-experimental 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 "texdata/sraka-27b-experimental" \
    --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": "texdata/sraka-27b-experimental",
		"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 "texdata/sraka-27b-experimental" \
        --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": "texdata/sraka-27b-experimental",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use texdata/sraka-27b-experimental with Ollama:
```
ollama run hf.co/texdata/sraka-27b-experimental:Q4_K_M
```
Unsloth Desktop

How to use texdata/sraka-27b-experimental with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf texdata/sraka-27b-experimental: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": "texdata/sraka-27b-experimental:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use texdata/sraka-27b-experimental with Docker Model Runner:
```
docker model run hf.co/texdata/sraka-27b-experimental:Q4_K_M
```

Lemonade

How to use texdata/sraka-27b-experimental with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull texdata/sraka-27b-experimental:Q4_K_M

Run and chat with the model

lemonade run user.sraka-27b-experimental-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use texdata/sraka-27b-experimental with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf texdata/sraka-27b-experimental: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 texdata/sraka-27b-experimental:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use texdata/sraka-27b-experimental with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf texdata/sraka-27b-experimental: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 "texdata/sraka-27b-experimental: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"

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Dostop do Srake je odobren ročno. Prosimo, navedite svoj e-naslov in namen uporabe. Access to Sraka is granted manually. Please provide your e-mail and intended use.

Sraka 27B

EKSPERIMENTALEN MODEL — samo za nekomercialno rabo. Raziskovalna izdaja: ni preizkušena v produkciji, ni varnostno poravnana in se lahko moti. Uporaba je dovoljena za raziskave, ocenjevanje in osebno rabo (CC BY-NC 4.0). Za komercialno rabo se obrnite na MediaAtlas Ltd.

EXPERIMENTAL MODEL — non-commercial use only. Research release: not production-tested, not safety-aligned, and it does get things wrong. Research, evaluation and personal use are permitted (CC BY-NC 4.0); contact MediaAtlas Ltd. for commercial licensing.

Slovenski splošnonamenski jezikovni model s 27 milijardami parametrov. Pogovor v slovenščini, sklepanje, prevajanje, klicanje orodij in daljše agentske naloge. Teče lokalno.

Avtor: Tadej Fius, MediaAtlas Ltd.
Osnova: Qwen3.8-27B (abliterated), dodatno naučen za slovenščino
Licenca: CC BY-NC 4.0 (nekomercialno). Osnovni model Qwen3.8-27B ostaja pod Apache-2.0 in za njegov delež veljajo njegovi pogoji; nekomercialna omejitev velja za našo uglasitev.
Status: eksperimentalen, raziskovalna izdaja
Kontekst: 262 144 tokenov; hibridna arhitektura (gated DeltaNet + polna pozornost vsak 4. sloj)
MTP glava: ohranjena (mtp.*, 15 tenzorjev) — omogoča spekulativno dekodiranje brez ločenega osnutkovnega modela. transformers te uteži pri nalaganju tiho zavrže, zato so bile po vsakem zlivanju LoRA ponovno vgrajene; brez tega bi model tekel počasneje, kot zna.
Vizualni stolp: ohranjen, med učenjem zamrznjen (enak osnovnemu modelu)

Sraka ni del družine GaMS. GaMS (Generative Model for Slovene) razvija skupina na Fakulteti za računalništvo in informatiko Univerze v Ljubljani v okviru programa PoVeJMo, objave pa tečejo pod okriljem CJVT; Sraka je samostojno delo podjetja MediaAtlas Ltd.

Rezultati

slovenian-llm-eval, n=500 na nalogo, 0-shot, acc_norm_char, generativno ocenjevanje:

naloga	osnova (Qwen3.8-27B abl.)	Sraka 27B
arc_easy	0.968	0.982
arc_challenge	0.942	0.950
hellaswag	0.686	0.752
piqa	0.852	0.862
openbookqa	0.876	0.920
winogrande	0.708	0.748
boolq	0.878	0.886
povprečje	0.844	0.871

Klicanje orodij (lasten held-out nabor naslednjega koraka, n=2471):

	osnova	Sraka 27B
odločitev klicati	0.893	0.922
izbira orodja	0.792	0.801
argumenti	0.628	0.651
pravilno ne klicati	0.847	0.956

Uporaba

vLLM

vllm serve MediaAtlas/sraka-27b --served-model-name sraka \
    --max-model-len 16384 --enable-auto-tool-choice \
    --tool-call-parser qwen3_xml --reasoning-parser qwen3

transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("MediaAtlas/sraka-27b")
model = AutoModelForCausalLM.from_pretrained("MediaAtlas/sraka-27b", dtype="bfloat16", device_map="auto")

msgs = [{"role": "user", "content": "Razloži razliko med dajalnikom in mestnikom."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=512)[0][ids.shape[1]:], skip_special_tokens=True))

Razmišljanje (<think>) je podprto; vklopi ga s enable_thinking=True v predlogi.

GGUF (llama.cpp / LM Studio / Ollama)

V mapi gguf/ so kvantizirane različice:

datoteka	velikost	za kaj
`sraka-27b-Q8_0.gguf`	~29 GB	referenca, praktično brez izgube
`sraka-27b-Q5_K_M.gguf`	~19 GB	najboljše razmerje kakovost/velikost
`sraka-27b-Q4_K_M.gguf`	~17 GB	24 GB kartica z daljšim kontekstom
`mmproj-sraka-27b-F16.gguf`	~1 GB	vizualni vhod (slike)
`mtp-sraka-27b-BF16.gguf`	~1 GB	MTP osnutkovna glava za spekulativno dekodiranje

llama-server -m sraka-27b-Q5_K_M.gguf --mmproj mmproj-sraka-27b-F16.gguf \
    -md mtp-sraka-27b-BF16.gguf --ctx-size 16384 --jinja

-md (osnutkovni model) je MTP glava: isti model, en korak naprej, zato pospeši generiranje brez spremembe izhoda. --jinja je potreben za klicanje orodij.

Omejitve

Eksperimentalen model. Raziskovalna izdaja brez produkcijskega testiranja; obnašanje se lahko med različicami spremeni. Nekomercialna raba (CC BY-NC 4.0).
Osnovni model je abliteriran (odstranjene zavrnitvene smeri) in Sraka to podeduje: nima zanesljivih varnostnih zavrnitev. Za javno rabo dodajte lastno moderacijo.
Znanje o Sloveniji je boljše od osnove, a ne popolno: še vedno se zmoti pri pripisovanju literarnih del (npr. Cvetje v jeseni pripiše Župančiču namesto Tavčarju). Preverjajte dejstva.
Ni namenjen samodejnemu odločanju o ljudeh (zaposlovanje, kredit, zdravje, pravo).
Vizualni stolp je med učenjem zamrznjen in je enak osnovnemu modelu; slikovne zmožnosti niso bile ne uglašene ne izmerjene.

Zasluge in viri podatkov

Qwen3.8-27B (Apache-2.0), abliteracija huihui-ai
Slovenska Wikipedija (CC BY-SA), cjvt/Wikipedia-translated (CC BY 4.0)
cjvt/sl_code_feedback (CC BY 4.0)
Sloleks 3.0 (CC BY-SA 4.0), CLARIN.SI
Orodja: APIGen-MT, ToolACE, Nemotron-SFT-Agentic-v2, Toucan-1.5M

Citiranje

@misc{sraka27b,
  title  = {Sraka 27B: a Slovenian general-purpose language model},
  author = {Fius, Tadej},
  year   = {2026},
  note   = {MediaAtlas Ltd.},
  url    = {https://huggingface.co/MediaAtlas/sraka-27b}
}