Instructions to use KucLab/kuclab-hertz-0.8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use KucLab/kuclab-hertz-0.8 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 KucLab/kuclab-hertz-0.8:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf KucLab/kuclab-hertz-0.8:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf KucLab/kuclab-hertz-0.8:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf KucLab/kuclab-hertz-0.8: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 KucLab/kuclab-hertz-0.8:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf KucLab/kuclab-hertz-0.8: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 KucLab/kuclab-hertz-0.8:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf KucLab/kuclab-hertz-0.8:Q4_K_M

Use Docker

docker model run hf.co/KucLab/kuclab-hertz-0.8:Q4_K_M

LM Studio
Jan

vLLM

How to use KucLab/kuclab-hertz-0.8 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "KucLab/kuclab-hertz-0.8"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "KucLab/kuclab-hertz-0.8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/KucLab/kuclab-hertz-0.8:Q4_K_M

Ollama
How to use KucLab/kuclab-hertz-0.8 with Ollama:
```
ollama run hf.co/KucLab/kuclab-hertz-0.8:Q4_K_M
```
Unsloth Desktop

How to use KucLab/kuclab-hertz-0.8 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf KucLab/kuclab-hertz-0.8: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": "KucLab/kuclab-hertz-0.8:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use KucLab/kuclab-hertz-0.8 with Docker Model Runner:
```
docker model run hf.co/KucLab/kuclab-hertz-0.8:Q4_K_M
```

Lemonade

How to use KucLab/kuclab-hertz-0.8 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull KucLab/kuclab-hertz-0.8:Q4_K_M

Run and chat with the model

lemonade run user.kuclab-hertz-0.8-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use KucLab/kuclab-hertz-0.8 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf KucLab/kuclab-hertz-0.8: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 KucLab/kuclab-hertz-0.8:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use KucLab/kuclab-hertz-0.8 with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf KucLab/kuclab-hertz-0.8: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 "KucLab/kuclab-hertz-0.8: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"

KucLab Hertz 0.8

Measured on this project's harness — same 240 MMLU-Pro STEM and 206 CZ terminology questions for every model.

A Czech/English STEM + programming assistant built by KucLab on top of Qwen/Qwen3.5-9B. Second-generation distillation: API teachers (DeepSeek-flash for STEM, Gemini 3.8 Flash for prose/Czech) with enforced answer-first discipline, trained with the proven Hertz recipe. It beats Hertz 0.7F on MMLU-Pro STEM.

What this is

Hertz 0.8 is a LoRA fine-tune (r=16, merged into the base weights) of Qwen3.5-9B:

Base: Qwen/Qwen3.5-9B (~9B params, Apache 2.0)
Method: QLoRA, r=16 / alpha=32, merged to bf16 then quantized
Teachers: DeepSeek-flash via API (math, physics, chemistry, programming — reasoning, temperature 0.2–0.3) and Gemini 3.8 Flash via API (biology, design, Czech language, concise everyday answers — thinking disabled for discipline)
Training data: ~5080 rows total — API-distilled STEM/prose/concise rows (answer-first, canonical Answer: (X) final line on multiple-choice), 309 CS↔EN scientific-terminology rows with definitions (train split only), 28 answer-first formatting examples, identity rows (Hertz 0.8). Validated: 0 malformed, 0 duplicates.
Context: 32768 tokens in the Ollama Modelfile (num_ctx).
Format available: GGUF (q4_k_m, ~5.3GB) for llama.cpp/Ollama.

Quickstart (Ollama)

Important: ollama pull hf.co/... alone does NOT apply this model's system prompt. Use ollama create with the Modelfile below instead:

curl -O https://huggingface.co/KucLab/kuclab-hertz-0.8/resolve/main/Modelfile
ollama create kuclab-hertz-0.8 -f Modelfile
ollama run kuclab-hertz-0.8

Development story

Hertz 0.7F (83.3% STEM) was distilled from a local 30B teacher. For 0.8 we moved teachers to API frontier models and fixed two data diseases found by audit: (1) ~370 truncated answers cut by token limits were removed — they teach broken patterns; (2) parametrized templates with absurd ranges (93kW kettles, 862 bpm heart rates) were constrained to realistic values. A brevity domain (400 short Q&A) teaches the model that "ahoj" deserves one sentence, not an essay — same spirit as ThinkingCap-style token efficiency. Qwen3.5 still reasons by architecture; append "think": false to API requests for instant answers.

Benchmarks

Same prompts, same grading code, same Ollama Q4_K_M quantization, identical methodology throughout.

MMLU-Pro STEM (240 held-out questions, this project's own curated subset)

	Hertz 0.6 (12B)	Hertz 0.7F (9B)	Hertz 0.8 (9B)
Biology	91.7%	83.3%	86.7%
Chemistry	61.7%	78.3%	83.3%
Math	90.0%	96.7%	95.0%
Physics	73.3%	75.0%	90.0%
Total	79.2%	83.3%	88.8%

Hertz 0.8 beats 0.7F by +5.5pp, with physics jumping +15pp. Only 10/240 answers needed the fallback re-ask (vs 22 in 0.7F).

Czech terminology benchmark (206 held-out CS↔EN scientific terms)

	Hertz 0.6	Hertz 0.7F	Hertz 0.8
CS→EN	82.5%	88.3%	87.4%
EN→CS	65.0%	82.5%	81.6%
Total	73.8%	85.4%	84.5%

Essentially tied with 0.7F (−0.9pp) — the API-teacher Czech prose held the gains.

Honest status

✅ MMLU-Pro STEM: 88.8%, beats Hertz 0.7F (83.3%)
✅ Physics 90.0%, chemistry 83.3% — the API STEM distillation worked
✅ Correctly identifies as KucLab Hertz 0.8 (kuclab.org), no founder named
⚠️ Czech terminology (84.5%) essentially tied with 0.7F, not above it
⚠️ 91% STEM target not yet reached (−2.2pp) — next iteration
⏳ No tool-calling fine-tuning (tool-use rows were cut with the budget)

License

Apache 2.0, inherited from Qwen/Qwen3.5-9B.

Credits

Base model: Qwen/Qwen3.5-9B (Qwen, Apache 2.0)
Teachers: DeepSeek-flash (STEM) and Gemini 3.8 Flash (prose/Czech) via API
Fine-tuning, dataset construction, and packaging: KucLab

Downloads last month: 13

GGUF

Model size

9B params

Architecture

qwen35

Hardware compatibility

4-bit

Model tree for KucLab/kuclab-hertz-0.8

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

Qwen/Qwen3.5-9B

Adapter

(668)

this model