Why everyone is building their own LLM
2024–2026 became the era of national language models. Governments around the world realised that depending on someone else's AI systems is a strategic risk. Data goes abroad, models don't understand the local context, and the infrastructure belongs to foreign companies.
Who has already built one
| Country | Project | Parameters | Budget |
|---|---|---|---|
| UAE | Falcon (TII) | 7B–180B | $50–200M |
| Saudi Arabia | ALLaM (HUMAIN) | 34B | Part of HUMAIN's $100B |
| France | Mistral AI | up to 123B | $3B+ raised |
| Turkey | Kumru AI (TUBITAK) | 14B | $10–50M |
| Kazakhstan | KAZ-LLM (ISSAI) | 8B + 70B | ~$5–20M |
Uzbekistan is the only large Central Asian country without its own language model. Meanwhile, in October 2024 the President signed an AI development strategy (PP-358), and the Ministry of Digital Technologies announced the development of a national model. The goal is $1.5 billion in AI products by 2030.
Who needs a custom LLM
Not only governments. Your own model solves specific business tasks that can't be covered through the OpenAI or Anthropic API.
Government
Data sovereignty. Citizens' requests, medical records, court cases — everything stays inside the country. Public services in the national language without depending on foreign APIs.
Banks and telecom
Compliance and security. Customer data doesn't go to OpenAI. Chatbots understand the local context: currency, legislation, products. Regulators require local storage.
IT companies
Specialised agents. A model trained on your APIs: Telegram, CRM, database. Tool calling without intermediate layers. Full control over behaviour.
Education
Adaptation to the local context. An AI tutor that knows the Uzbek school curriculum, explains in the native language and takes cultural specifics into account.
Startups
80–90% savings. GPT-4o costs $2.50–15 per million tokens. Your own 14B model on one GPU — $0.30–0.50/1M. At scale, that's a difference of tens of thousands of dollars.
Content and media
Generation in the national language. GPT-4o and Claude write Uzbek mediocrely. Your own model means correct grammar, style and cultural code.
When you DON'T need your own model
An honest analysis. Your own LLM isn't a silver bullet. Here's when it's better to use ready-made APIs:
You need GPT-4o / Claude level
If the task requires frontier quality — complex reasoning, senior-level code generation, multimodality — a 14B model won't replace a 200B+ model. Use the API and RAG.
No training data
Fine-tuning without good data is useless. You need at least 500–1,000 labelled examples for a basic result, 5,000+ for a good one. If there's no data — start with prompt engineering.
Budget under $500
Fine-tuning itself costs $2–8 per run. But data preparation, testing, iterations and infrastructure add up to $500–1,200 for a prototype. If the budget is tightly limited — use the API with optimisation.
Rule: start with an API (Anthropic, OpenAI) + RAG. If you hit a wall on cost, privacy or quality in your language — then fine-tune.
What we built · MILA LLM v0.1
We went all the way from idea to a working model. Here's a step-by-step breakdown of our case.
Qwen3-14B
Why this model: Apache 2.0 licence (complete freedom), native support for 119 languages including Uzbek, 14 billion parameters — enough for good answers, and it fits on one GPU.
QLoRA Fine-Tuning
We freeze all 8.8 billion parameters and compress them to 4-bit. We add small trainable adapters — 256M parameters (2.91% of the total). We train only those. Result: the model learns new things without forgetting the old.
12 custom examples
For the proof of concept: 8 function-calling examples (Telegram, Calendar, Linear, Convex), 2 examples of Uzbek knowledge, 2 examples of domain knowledge. ChatML format. Language mix: Uzbek (Latin script), Russian, English.
RunPod A100, ~20 minutes
NVIDIA A100 80GB on RunPod. The full cycle: creating the pod, uploading the data, training (18 seconds on 12 examples), merging the LoRA adapters, GGUF quantisation, downloading the result. Total: ~$15.
Result
| Parameter | Value |
|---|---|
| GGUF size | 8.4 GB (Q4_K_M) |
| Inference | ollama run mila-llm:14b |
| Languages | Uzbek, Russian, English |
| Function calling | Telegram, Calendar, Linear, Web |
| Cost | ~$15 for a full run |
The model runs locally on an ordinary Mac via Ollama. It answers in Uzbek, Russian and English. It calls tools via tool_calls. The full pipeline is debugged and reproducible.
Economics · Custom LLMs
Cost comparison
| GPT-4o API | Fine-tuned 14B (self-hosted) | Fine-tuned 14B (cloud) | |
|---|---|---|---|
| Cost | $2.50–15/1M tokens | $0 (hardware costs) | $0.30–0.50/1M |
| Data | Goes to OpenAI | Stays with you | On your server |
| Customisation | Prompt only | Full | Full |
| Latency | ~500 ms | ~100 ms (GPU) | ~200 ms |
Training cost
- QLoRA (we recommend it): updates 0.1% of parameters, 10–16 GB VRAM — $2–8 per run
- LoRA: updates 0.1–1% of parameters, 24–48 GB VRAM — $50–200
- Full fine-tuning: all parameters, ~200 GB VRAM — $2,000–5,000
How many users one GPU can serve
| GPU | Purchase | Rent/hour | Users |
|---|---|---|---|
| RTX 4090 | $1,600–2,000 | $0.34–0.44 | 30–60 |
| L40S | $7,000–10,000 | $0.59–1.24 | 50–100 |
| A100 80GB | $10,000–15,000 | $0.67–2.06 | 80–150 |
| H100 SXM | $27,000–40,000 | $1.99–6.16 | 150–300 |
One RTX 4090 for $1,600 serves 30–60 users at the same time. For a startup or a pilot project, one card is enough.
Roadmap · From prototype to product
Prototype
MILA-LLM-14B v0.1 based on Qwen3-14B. QLoRA, 12 examples, GGUF 8.4 GB, runs via Ollama. Cost: ~$15.
Validation
5,000–10,000 training examples. Benchmarks: FC accuracy, Uzbek QA, MMLU. A/B testing against the base model. Publication on HuggingFace. First pilot clients.
Scaling
Full fine-tuning on a government GPU cluster (NVIDIA Blackwell B200). An expanded language corpus. DPO alignment. Model variants: 4B (edge), 14B (standard), 32B (premium).
Production
4–8 GPUs in Tashkent, an inference API, 10+ clients. A UzbekMMLU benchmark. Partnership with the Ministry of Digital Technologies. Inference as a service for business.
What's needed for the next stage
- A GPU cluster — for full fine-tuning and inference (available through a government programme or a grant)
- A language corpus — Uzbek text: news, books, documents (a Ministry of Digital Technologies project)
- A $15–40K grant — for ML engineers, data expansion, infrastructure
MILA LLM — an open-source model for Uzbekistan
A working prototype. A complete pipeline. Apache 2.0 licence.
Open to partnership: government, business, investors, ML engineers.