Instructions to use Kaidris/Kaidris-ABAP-7B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Kaidris/Kaidris-ABAP-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 Kaidris/Kaidris-ABAP-7B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Kaidris/Kaidris-ABAP-7B:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Kaidris/Kaidris-ABAP-7B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Kaidris/Kaidris-ABAP-7B: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 Kaidris/Kaidris-ABAP-7B:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Kaidris/Kaidris-ABAP-7B: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 Kaidris/Kaidris-ABAP-7B:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Kaidris/Kaidris-ABAP-7B:Q4_K_M

Use Docker

docker model run hf.co/Kaidris/Kaidris-ABAP-7B:Q4_K_M

LM Studio
Jan

vLLM

How to use Kaidris/Kaidris-ABAP-7B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Kaidris/Kaidris-ABAP-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": "Kaidris/Kaidris-ABAP-7B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Kaidris/Kaidris-ABAP-7B:Q4_K_M

Ollama
How to use Kaidris/Kaidris-ABAP-7B with Ollama:
```
ollama run hf.co/Kaidris/Kaidris-ABAP-7B:Q4_K_M
```
Unsloth Desktop

How to use Kaidris/Kaidris-ABAP-7B with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Kaidris/Kaidris-ABAP-7B: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": "Kaidris/Kaidris-ABAP-7B:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Kaidris/Kaidris-ABAP-7B with Docker Model Runner:
```
docker model run hf.co/Kaidris/Kaidris-ABAP-7B:Q4_K_M
```

Lemonade

How to use Kaidris/Kaidris-ABAP-7B with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Kaidris/Kaidris-ABAP-7B:Q4_K_M

Run and chat with the model

lemonade run user.Kaidris-ABAP-7B-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use Kaidris/Kaidris-ABAP-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 Kaidris/Kaidris-ABAP-7B: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 Kaidris/Kaidris-ABAP-7B:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Kaidris/Kaidris-ABAP-7B with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Kaidris/Kaidris-ABAP-7B: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 "Kaidris/Kaidris-ABAP-7B: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"

Kaidris-ABAP-7B — Community Edition

A small, local ABAP coding model: modern ABAP (7.50+) classes, ABAP SQL, internal-table logic and ABAP Unit-testable code. It runs from a 6 GB GPU (Q4) or on CPU only, and nothing leaves your machine.

Free: Apache-2.0, commercial use included.

Coming next: Kaidris Forge. Our paid model, now in development. It will be released only once it scores at least twice this model on our hardest exams: multi-method business apps and problems needing techniques it was never shown. The numbers will be published with the release. It adds a write → compile → test → fix agent and a dedicated repair model. Planned launch pricing: Individual $1,000 / year (self-paid, personal use) · Enterprise $2,000 / user / year. Join the waitlist: kaidris.com/sap/forge

Results

pass@1, greedy, one answer per task. Every answer is compiled and its ABAP Unit tests are run by an open-source ABAP runtime (abaplint transpiler + open-abap), plus our own strictness lint. These are not scores from an SAP system; see How it was measured.

Model	General ABAP benchmark (158 tasks)	ABAP SQL, unseen schemas (165)	ABAP SQL Dev (44)	ABAP SQL Exam A (44)
Qwen2.5-Coder-7B-Instruct (base)	22	53	13	12
Qwen3.5-9B	17	—	—	—
Seed-Coder-8B-Instruct	18	—	—	—
oisee abap-v6 (7B)	24	46	9	13
Kaidris-ABAP-7B (this model, Q8_0)	51	149 (90%)	35	38

General ABAP benchmark: 158 HumanEval-style ABAP tasks (one class per task, checked by unit tests).
ABAP SQL, unseen schemas: 165 tasks on two business schemas this model never saw in training.
Dev and Exam A: held-out tasks on the training schemas.

Business-app exam (4 classes, 16 methods, 146 tests on unseen schemas): one shot, 8/16 methods right.

Limitations, said plainly

New techniques. Recursion, date arithmetic, remainder-exact allocation and similar are weak. On our Novel exam (9 module scenarios) it is no better than the base model. That is what Kaidris Forge is for.
Multi-method classes. Whole classes in one answer often have one wrong method. Ask for one method at a time.
Not checked on SAP. It has not been validated on a licensed SAP system. Always run your own ATC checks and unit tests.
SQL dialect. It writes ABAP SQL for classic tables; CDS / RAP are not covered yet.

Files

File	Size	Use	ABAP SQL, unseen schemas
`kaidris-abap-7b-Q8_0.gguf`	7.5 GB	best quality, 10 GB+ GPU	149/165
`kaidris-abap-7b-Q5_K_M.gguf`	5.1 GB	8 GB GPU	147/165
`kaidris-abap-7b-Q4_K_M.gguf`	4.4 GB	laptops / CPU	147/165

Minimum requirements

Memory needed = model + context cache + working buffers. Measured with llama.cpp at an 8,192-token context; the context cache grows linearly, about +0.45 GB per extra 8k tokens.

File	Memory needed (8k context)	GPU (VRAM)	Without a GPU (CPU only)
Q4_K_M	≈ 5.1 GB	6 GB minimum, 8 GB recommended	8 GB RAM minimum, 16 GB recommended (slower)
Q5_K_M	≈ 5.9 GB	8 GB minimum	12 GB RAM minimum, 16 GB recommended
Q8_0	≈ 8.4 GB	10 GB minimum, 12–16 GB recommended	16 GB RAM minimum

GPU: NVIDIA (CUDA), AMD (ROCm / Vulkan) or Apple Silicon (Metal) via llama.cpp. Leave about 1 GB of VRAM for the desktop.
Disk: the file size + about 1 GB.
OS: Windows 10/11, Linux or macOS. Tested on Windows 11 with an RTX 5080.
Speed: depends heavily on the hardware and was not benchmarked for this card. A GPU is much faster than CPU only.

Works with llama.cpp, LM Studio and Ollama. Example:

llama-server -m kaidris-abap-7b-Q8_0.gguf -ngl 99 -c 8192 --jinja

Prompt format (what it was trained on)

System:

You are an expert SAP ABAP developer. Write modern ABAP (7.50+). Answer with one complete ABAP class in a single ```abap code block. No explanations.

User:

Task:
Return how many rows of ZKD_ORDER have STATUS = iv_value (0 if none).
Table columns:
ZKD_ORDER: ORDER_ID (NUMC 8, key), CUST_ID (NUMC 6), ORDER_DATE (DATS), STATUS (CHAR 1), ...

Implement it in ABAP using exactly this class definition:
```abap
CLASS zcl_demo DEFINITION PUBLIC FINAL CREATE PUBLIC.
  PUBLIC SECTION.
    CLASS-METHODS count_by_status
      IMPORTING iv_value TYPE zkd_order-status RETURNING VALUE(rv_count) TYPE i.
ENDCLASS.
```

Giving the table columns and the exact class definition matters most for good results.

How it was made (data provenance)

No SAP software, no SAP system output and no closed-model output were used to train this model.

Base: Qwen2.5-Coder-7B-Instruct (Apache-2.0), QLoRA fine-tune.
General ABAP (708 examples): problems from MBPP (CC-BY-4.0), LeetCode problems via the newfacade/LeetCodeDataset dataset (Apache-2.0 on Hugging Face, MIT on GitHub; the original problem statements are LeetCode's, adapted into ABAP tasks) and self-oss-instruct (ODC-BY); solutions written by Qwen3.8-27B (Apache-2.0); each one kept only if it compiled and passed its unit tests on the open-source runtime.
ABAP SQL (279 verified solutions): tasks on invented business tables (sales, inventory, HR, projects, billing, maintenance), written by Qwen3.8-27B, and kept only if every test passed on the open-source runtime (SQLite-backed) and our lint.
Evaluation tables: the two exam schemas, the business-app exam and the Novel exam were never used for training.
Tooling: the training and evaluation tooling was built with Claude Code (Anthropic). This model learned only from code written by an open Apache-2.0 model (Qwen3.8-27B), each example verified by an open-source ABAP runtime.

How it was measured

Runtime: the open-source runtime (abaplint transpiler and open-abap, MIT) executes ABAP and ABAP Unit; ABAP SQL runs on SQLite.
Known differences: it is close to a real system but not identical. We found 8 constructs where it rejects valid ABAP and documented each one. A lint pass adds checks it does not enforce (e.g. @ escaping, UP TO after INTO, 255-character lines).
Controls: every exam was validated first. Reference solutions must pass, empty or single-method-gutted solutions must fail, and planted subtle bugs must be caught.

Licence and attribution

Licence: Apache-2.0. Kaidris-ABAP-7B © 2026 Kaidris LLC. The base model is Qwen2.5-Coder-7B-Instruct © Alibaba Cloud (Apache-2.0).
Training-problem sources: MBPP (CC-BY-4.0, Google Research), self-oss-instruct (ODC-BY, BigCode).
Trademarks: SAP and ABAP are trademarks of SAP SE. This model is not affiliated with, endorsed by or sponsored by SAP.

Built by Kaidris

Kaidris builds small, local, verifiable AI models. More at kaidris.com/sap.

Downloads last month: 120

GGUF

Model size

8B params

Architecture

qwen2

Hardware compatibility

4-bit

5-bit

8-bit

Model tree for Kaidris/Kaidris-ABAP-7B

Base model

Qwen/Qwen2.5-7B

Finetuned

Qwen/Qwen2.5-Coder-7B

Finetuned

Qwen/Qwen2.5-Coder-7B-Instruct

Quantized

(268)

this model