LLM inference engine

[stvowi]

PasLLM is a LLM inference engine in Object Pascal.

This makes it possible to create Unleashed Pascal programmes with AI capabilities without the usual need to install endless dependencies.
I therefore propose that PasLLM be pre-installed in the IDE.

Hey everyone, please give this project as many stars as possible to encourage the developer to keep working on it.

Unfortunately, it’s only a half-finished project.

@see also:

• What is LLM inference?
• PasLLM - AI LLM inference engine in Object Pascal (1)
• PasLLM - AI LLM inference engine in Object Pascal (2)
• Lazerus Forum - PasLLM - LLM Inference Engine in Pure Pascal
• Deutschsprachiges Forum zur Entwicklungsumgebung Lazarus - PasLLM & verschiedene Modelle

The PasLLM project appears to have been put on hold and is not yet ready for production.
It seems one can only make use of what is already available.

[stvowi]

You can download PasLLM models from:

• Pre-converted and quantized models: [PasLLM model folder on MEGA](https://mega.nz/folder/krcgHCpZ#0tjLqup_Hc4THWC9itDrTg)
• PasLLM and conversion tools: [BeRo1985/pasllm on GitHub](https://github.com/BeRo1985/pasllm)
• Original models: [Hugging Face](https://huggingface.co/models)

Place downloaded PasLLM models in:

pasllm/bin/models/

Example:

./bin/pasllmcli \
  -model=bin/models/qwen2.5_0.5b_instruct_q40nl.safetensors

PasLLM does not directly load arbitrary GGUF files. It uses its own Safetensors-based format with quantization types such as Q40NL, Q40, Q80, FP8, and FP16. Models from Hugging Face must be converted using tools/convert.py, as documented in the [PasLLM README](https://github.com/BeRo1985/pasllm).

Good models for initial testing include:

• Qwen 2.5 0.5B Instruct
• Qwen 3 0.6B
• SmolLM2 360M Instruct
• TinyLlama 1.1B Chat

Remember to check the license of each original Hugging Face model.

[stvowi]

For Lazarus/Free Pascal development, _the best practical model is Qwen3-Coder 30B—especially if you want a capable local assistant through Ollama. It is code-specialized, supports a 256K context window, and its MoE design activates only 3.3B parameters per request, making it unusually efficient for its class. ollama - qwen3-coder
For the highest code quality on difficult multi-unit refactors, debugging, and architecture, use a frontier cloud coding model alongside it; but for your 32 GB RAM / 12 GB VRAM local workflow, Qwen3-Coder 30B is the best starting point.

Recommended setup

Use case	Recommended model	Why
Best local coding assistant	qwen3-coder:30b	Purpose-built for agentic software engineering and coding; 19 GB download with 256K context. [ollama]
Larger local/server setup	qwen3-coder-next	Coding-focused 80B MoE model, but the Q4 build alone is 52 GB, so it is not a good fit for your present machine. [ollama]
General Pascal reasoning	qwen3:14b	A useful lighter fallback for explanations, API design, and reviewing compact units. [ollama]
Maximum quality, cloud	Qwen3-Coder 480B Cloud	Stronger for repository-scale work, but the local model requires about 290 GB. [ollama]

Why this fits Pascal

No leading model is specifically trained or benchmarked primarily on Free Pascal/Lazarus,
so expect occasional Delphi-centric code, incompatible syntax, or assumptions about VCL rather than LCL. The decisive advantage comes from supplying the model with your target details: FPC version, Lazarus/LCL widgetset, operating system, compiler switches, and the involved unit interfaces.
Qwen3-Coder is explicitly positioned for long-context, agentic coding workflows and can be launched through Ollama with coding-agent tooling. That is particularly useful for your likely tasks: tracing dependencies across units, improving PostgreSQL access code, generating tests, and diagnosing compiler messages. ollama - qwen3-coder

Start locally

ollama pull qwen3-coder:30b
ollama run qwen3-coder:30b

The official Ollama catalogue lists this 30B model at approximately 19 GB and with a 256K context window. ollama - qwen3-coder
If responsiveness or Intel Arc support becomes a limiting factor, use:

ollama pull qwen3:14b
ollama run qwen3:14b

It will be less reliable on multi-file refactoring, but is a sensible interactive fallback. ollama - qwen3:14b

Prompt template

Use a stable project instruction such as:

You are an expert Free Pascal and Lazarus developer.
Target:

Free Pascal: 3.2.x
Lazarus: 3.x
GUI: LCL, not Delphi VCL
Platforms: Windows, Linux, macOS
Database: PostgreSQL via [specify library]
Compatibility: avoid Delphi-only language and RTL features

Rules:

Return compilable Object Pascal.
State required units in the uses clauses.
Preserve existing public interfaces unless explicitly asked to change them.
Prefer LCL and FPC-compatible APIs.
Identify ownership, lifetime, thread-safety, exception handling, and SQL-parameter issues.
If uncertain about an API, say so instead of inventing it.

Task:

[paste compiler error, unit interface, implementation, and expected behavior]

For best results, give it a small, coherent slice of the codebase: the failing unit, related interfaces, exact compiler output, and expected behavior. Ask it first to analyse and propose a patch, then ask it to emit a complete revised unit; this reduces accidental API changes.

For Lazarus/Free Pascal development, the best practical model is Qwen3-Coder 30B—especially if you want a capable local assistant through Ollama. It is code-specialized, supports a 256K context window, and its MoE design activates only 3.3B parameters per request, making it unusually efficient for its class. [ollama](https://ollama.com/library/qwen3-coder)

For the highest code quality on difficult multi-unit refactors, debugging, and architecture, use a frontier cloud coding model alongside it; but for your 32 GB RAM / 12 GB VRAM local workflow, Qwen3-Coder 30B is the best starting point.

Recommended setup

Use case	Recommended model	Why
Best local coding assistant	qwen3-coder:30b	Purpose-built for agentic software engineering and coding; 19 GB download with 256K context. ollama](https://ollama.com/library/qwen3-coder)
Larger local/server setup	qwen3-coder-next	Coding-focused 80B MoE model, but the Q4 build alone is 52 GB, so it is not a good fit for your present machine. ollama](https://ollama.com/library/qwen3-coder-next)
General Pascal reasoning	qwen3:14b	A useful lighter fallback for explanations, API design, and reviewing compact units. [ollama](https://ollama.com/library/qwen3:14b)
Maximum quality, cloud	Qwen3-Coder 480B Cloud	Stronger for repository-scale work, but the local model requires about 290 GB. [ollama](https://ollama.com/library/qwen3-coder)

Why this fits Pascal

No leading model is specifically trained or benchmarked primarily on Free Pascal/Lazarus, so expect occasional Delphi-centric code, incompatible syntax, or assumptions about VCL rather than LCL. The decisive advantage comes from supplying the model with your target details: FPC version, Lazarus/LCL widgetset, operating system, compiler switches, and the involved unit interfaces.

Qwen3-Coder is explicitly positioned for long-context, agentic coding workflows and can be launched through Ollama with coding-agent tooling. That is particularly useful for your likely tasks: tracing dependencies across units, improving PostgreSQL access code, generating tests, and diagnosing compiler messages. [ollama](https://ollama.com/library/qwen3-coder)

Start locally

ollama pull qwen3-coder:30b
ollama run qwen3-coder:30b

The official Ollama catalogue lists this 30B model at approximately 19 GB and with a 256K context window. [ollama](https://ollama.com/library/qwen3-coder)

If responsiveness or Intel Arc support becomes a limiting factor, use:

ollama pull qwen3:14b
ollama run qwen3:14b

It will be less reliable on multi-file refactoring, but is a sensible interactive fallback. [ollama](https://ollama.com/library/qwen3:14b)

Prompt template

Use a stable project instruction such as:

You are an expert Free Pascal and Lazarus developer.

Target:
- Free Pascal: 3.2.x
- Lazarus: 3.x
- GUI: LCL, not Delphi VCL
- Platforms: Windows, Linux, macOS
- Database: PostgreSQL via [specify library]
- Compatibility: avoid Delphi-only language and RTL features

Rules:
- Return compilable Object Pascal.
- State required units in the uses clauses.
- Preserve existing public interfaces unless explicitly asked to change them.
- Prefer LCL and FPC-compatible APIs.
- Identify ownership, lifetime, thread-safety, exception handling, and SQL-parameter issues.
- If uncertain about an API, say so instead of inventing it.

Task:
[paste compiler error, unit interface, implementation, and expected behavior]

For best results, give it a small, coherent slice of the codebase: the failing unit, related interfaces, exact compiler output, and expected behavior. Ask it first to analyse and propose a patch, then ask it to emit a complete revised unit; this reduces accidental API changes.

[stvowi]

AI Prompt Engineering Course - Prompt Engineering Beginner COMPLETE Guide and for PROS (2025)

[stvowi]

Sorry, there is a error in the text, Install Qwen with: ollama run qwen3-coder !

[stvowi]

If you want play with PasLLM-Models you can download a Windows-App .

[stvowi]

By the way, the preferred language for prompts is English. Apparently, this gives the best results.

Have fun!

[stvowi]

There’s a lovely video on the subject of ‘models’:
The Best Local AI To Run On 8GB VRAM

[stvowi]

If a model isn’t up to scratch for your purposes, you’ll need to train it using your specialist knowledge.
Unslot now offers a simple way to improve a model, provided you have the right hardware hardware .

@see also:

• First impressions
• Unsloth's Local Deep Research: Is it any good?

[stvowi]

A tool as IDE-Addon such like Unslot with/ for pasLLM would be great.

The IDE needs unique features to attract attention and gain popularity.
You can only learn / test Unleashed Pascal with a real project.

[stvowi]

Well, have fun with SLM!

@see also:

• Was ich über lokale KI gelernt habe
• This Tiny Model Changes How We Think About Local AI
• This tiny LLM dominates RAG and is SUPER FAST
• Great Small LLMs for the GPU Poor Like Me!
• Cactus Needle - The 26M Function Calling Model
• Needle: Finetune a 26M Tool-Calling Model Locally with Ollama
• Gemma 4 Coder: 12B Model Carrying Fable 5's Reasoning on 8GB VRAM, Fully Offline
• Introducing Ornith 1.0 - Agentic Coding LLMs
• MiniCPM5-2B: The Best Sub-Agent Model Yet?

[stvowi]

How to make your local AI 10 times more powerful...

[stvowi]

The Best Local AI To Run On 16GB VRAM

[stvowi]

Frontier-Class AI at Home. Is It Finally Here?

[stvowi]

...that is crazy: Build the Gemma Translator: No Internet Required

[stvowi]

The BEST local AI music generator is here!

[stvowi]

MiniMax 3 Just Killed Suno - Everything You Need To Know

[stvowi]

Google Gemma 4 12B: The End of Cloud-Only AI?

[stvowi]

Let’s turn Pascal’s source code into an audiobook in a IDE - Addon!?
Pascal’s syntax is bound to sound better than that of other programming languages.

• BreezeTTS2 - 100% Local Real-Time Voice
• Someone Built Ollama for Audio… and It Runs Locally
• Developers Might Finally Have a Local TTS Model That Doesn’t Suck

[stvowi]

How-To Create Free Local Voice AI with Pipecat + Ollama + Kokoro

[fibodevy]

If you're looking for people interested in Free Pascal, the official forum is the place to post: https://forum.lazarus.freepascal.org/

That's where the FPC/Lazarus crowd actually is, so you'd get a much bigger response there. It even has a board dedicated to AI in Pascal, which is exactly where PasLLM, local models and tooling recommendations belong. Far more readers and far more replies than anywhere else.

[stvowi]

Okay - I respect your wish.
I thought you wanted a few ideas.

Have a good one.

[fibodevy]

That's not what I meant at all - ideas are welcome here, keep them coming.

My point was only that you're posting a lot and getting almost no discussion back. The forum has far more people, so a thread there actually turns into a conversation. I read and post there myself, so it's not like I'd miss it.

You can keep posting here, that's fine 😃 I just don't have much time for watching videos - I've been fixing compiler bugs since this morning. Funnily enough it's AI related: I'm trying to fine-tune a small LLM to write Unleashed Pascal first, and fall back to classic FPC when it can't.

[stvowi]

The videos are for your followers, newbies and uneducaded visitors.

[stvowi]

I like it that we can fall back to classic FPC. This allows a smooth migration of legacy code to Unleashed Pascal.

[stvowi]

Please focus on your roadmap.

[stvowi]

Why You Should Stop Paying For AI (Local AI)

[stvowi]

LLM Quantization Explained: The Q4 Quality Trap

[stvowi]

Python vs Free Pascal (simple speed comparison)

[stvowi]

The Truth About AGI: Are We Already Past the Point of No Return?