A free model that writes modern ABAP on your own machine — nothing leaves it. More than twice its base model on general ABAP, 90% on unseen-schema ABAP SQL. Kaidris Forge, in development, adds an agent that compiles, tests and fixes the code before you see it.
Illustration: the prompt format from the model card and an example answer. Measured results below.
Modern ABAP 7.50+ABAP SQLABAP UnitRuns offlineFrom a 6 GB GPU — or CPU onlyApache-2.0No SAP system output in trainingNo closed-model outputllama.cpp · LM Studio · OllamaVerified by an open-source ABAP runtimeModern ABAP 7.50+ABAP SQLABAP UnitRuns offlineFrom a 6 GB GPU — or CPU onlyApache-2.0No SAP system output in trainingNo closed-model outputllama.cpp · LM Studio · OllamaVerified by an open-source ABAP runtime
01 Why it is different
Small, local, and checked line by line.
Runs on your machine
From a 6 GB GPU, or CPU only with the Q4 file and 8 GB of RAM. Works with llama.cpp, LM Studio and Ollama.
Your code, prompts and table definitions never leave the machine.
Every example passed its tests
Each training example was kept only if it compiled and passed its ABAP Unit tests on an open-source ABAP runtime.
987 verified examples: 708 general ABAP, 279 ABAP SQL on invented business tables.
No SAP output, no closed models
No SAP software, no SAP system output and no closed-model output were used to train it.
Every training solution was written by an open Apache-2.0 model (Qwen3.8-27B).
02 Results
More than twice the base model on general ABAP.
Measured, not promised. Pass@1, greedy, one answer per task — every answer compiled and unit-tested.
51of 158 general ABAP tasksbase model: 22
149of 165 ABAP SQL tasks on unseen schemasbase model: 53
4.4 GBQ4 file that still solves 147 of those 165runs on a laptop or CPU
Scored on an open-source ABAP runtime (abaplint transpiler + open-abap), not on an SAP system.
Show the full table
Tasks solved, pass@1, greedy decoding
Model
General ABAP 158 tasks
ABAP SQL, unseen schemas 165 tasks
ABAP SQL Dev 44 tasks
ABAP SQL Exam A 44 tasks
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 (Q8_0)
51
149 (90%)
35
38
Scored on an open-source ABAP runtime (abaplint transpiler + open-abap), not on an SAP system. Tasks solved, pass@1, greedy, one answer per task; every answer is compiled and its ABAP Unit tests are run, plus a strictness lint. Source: the model card (opens in new tab).
Limitations, said plainly
New techniques — recursion, date arithmetic, remainder-exact allocation and similar are weak. On our Novel exam it is no better than the base model. That is what Forge is for.
Multi-method classes — whole classes in one answer often have one wrong method; on the business-app exam it got 8 of 16 methods right in one shot. Ask for one method at a time.
Not checked on SAP — it has not been validated on a licensed SAP system. Run your own ATC checks and unit tests.
SQL dialect — ABAP SQL for classic tables; CDS and RAP are not covered yet.
03 Minimum requirements
Runs on the machine you already have.
Memory needed = model + context cache + working buffers, measured with llama.cpp at an 8,192-token context.
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
Longer context: the context cache grows linearly, about +0.45 GB per extra 8k tokens.
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 has not been benchmarked. A GPU is much faster than CPU only.
04 Private by design
Your ABAP never leaves the building.
The model, the prompts and your table definitions stay on your machine.
Works fully offline — on a plane, behind a firewall, on-premise.
No telemetry in the free model. It sends nothing, ever.
Forge adds one small licence check-in: a key hash, a machine-fingerprint hash and the app version. Never code.
05 How it works
Write, compile, test, fix — on your machine.
The free model writes the code. Forge wraps it in an agent that runs the whole loop locally and hands you code that already passes its tests.
Free model — available nowForge agent — in development
How it was builtThe 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.
Want it trained on your own codebase? Kaidris Custom — planned, from $15,000 setup.
07 FAQ
Questions, answered plainly.
Does my code leave my machine?
No. The free model runs entirely on your machine and sends nothing anywhere. Forge will add one network call: a periodic licence check-in that sends a hash of your licence key, a hash of a machine fingerprint and the app version — never code, prompts or file names.
Which ABAP versions does it cover?
Modern ABAP 7.50+: classes, ABAP SQL on classic tables, internal-table logic and ABAP Unit-testable code. CDS and RAP are not covered yet.
Is it tested on SAP?
Scores come from an open-source ABAP runtime; validation on licensed SAP systems is done in customer pilots. Always run your own ATC checks and unit tests on generated code.
What hardware do I need?
From a 6 GB GPU (8 GB recommended) for the Q4_K_M file, or CPU only with 8 GB of RAM (slower). The best-quality Q8_0 file needs a 10 GB GPU (12–16 GB recommended) or 16 GB of RAM. NVIDIA, AMD and Apple Silicon all work through llama.cpp, LM Studio or Ollama. Full table under Minimum requirements.
How is it different from GitHub Copilot or SAP Joule?
It is specialised in ABAP, it runs offline on your own hardware, and Forge adds an agent that compiles and unit-tests its own code and repairs what fails before you see it.
Can I use the free model commercially?
Yes. Kaidris-ABAP-7B Community Edition is Apache-2.0, commercial use included.
When does Forge ship, and what will it cost?
Forge is released only once it scores at least twice the free model on our hardest exams; the numbers are published with the release. Planned launch pricing: $1,000 per year for individuals and $2,000 per user per year for companies — see Pricing.
Refunds, offline use, new machines, lapsed licences?
Forge licences are planned for one named user on up to two machines, with self-service deactivation to move to a new PC, and an offline grace period between check-ins. The refund window and what happens when a licence lapses will be set out in the Forge licence agreement and refund policy, published before Forge goes on sale.