Berk ErdoganIndependent developer for quantized LLMs, local inference, and native apps
Based in Istanbul. Open-source model builds on Hugging Face, a self-hosted one-million-token inference server, and apps on the App Store.
Select a block on the die to zoom in. Esc zooms out.Tap a block on the die to zoom in.
Services
I am an independent developer in Istanbul. I quantize open-weight language models for NVIDIA Blackwell GPUs and publish them on Hugging Face, where they have more than 58,000 downloads. I build and run my own inference server, and I ship native apps to the App Store.
- P0 · BE-2026-APP
iOS and Android apps
Native SwiftUI and Kotlin apps, from the first screen to App Store release: StoreKit 2 subscriptions, App Store Connect, and Search Ads.
- P1 · BE-2026-LLM
LLM quantization
NVFP4, FP8, and INT8 builds of open-weight models, with calibration-set design and KL / ΔNLL evaluation against the BF16 source.
- P2 · BE-2026-LLM
Private LLM inference
Local GPU servers with vLLM: multi-GPU serving, long context, FP8 KV cache, and API endpoints for your own tools.
- P3 · BE-2026-WEB
Websites and business software
TypeScript sites on Cloudflare, and desktop business apps in SwiftUI or Electron.
Proof
- 5 open-source quantized LLMs, 58,844 all-time downloads
- 1,000,000-token context validated on 2× RTX 5090: 4/4 needle retrievals at 989k tokens
- 2 apps released on the App Store, with subscriptions and paid acquisition
- Business software in daily use at a 5-employee firm
Models
| Model | Base model | Format | Toolchain | Released | Downloads |
|---|---|---|---|---|---|
| Qwen3.6-27B-NVFP4 | Qwen3.6-27B (vision-language) | NVFP4 W4A4 | NVIDIA ModelOpt | 2026-05-02 | 49,909 |
| Qwen3.8-27B-NVFP4-1M | Qwen3.8-27B | NVFP4 + FP8, FP8 KV cache | llm-compressor, GPTQ | 2026-08-20 | 562 |
| Qwen3.8-27B INT8 AutoRound (community variant) | Qwen3.8-27B community variant | INT8 W8A16, group 128 | AutoRound (SignRound) | 2026-08-23 | 1,310 |
| gemma-4-12B-it-NVFP4 | gemma-4-12B-it (multimodal) | NVFP4 | NVIDIA ModelOpt | 2026-06-04 | 3,978 |
| Qwen3.5-27B-NVFP4 | Qwen3.5-27B (multimodal) | NVFP4 | NVIDIA ModelOpt | 2026-03-10 | 3,085 |
| Total | 58,844 |
The plot is for wider screens. The same values are in the table above.
GPU 6–10 · reserved for fine-tuned models
- GPU 6Reserved for a fine-tune
- GPU 7Reserved for a fine-tune
- GPU 8Reserved for a fine-tune
- GPU 9Reserved for a fine-tune
- GPU 10Reserved for a fine-tune
Fine-tuned models are in preparation. They will appear here and on huggingface.co/berkerdooo when the weights are published.
| Parameter | Value |
|---|---|
| Context length, validated (1) | 1,000,000 tokens |
| GPU memory, 2× RTX 5090 | 64 GB |
| System memory, DDR5 | 128 GB |
| Tensor parallel degree | 2 GPUs |
| Interconnect | PCIe |
(1) Qwen3.8-27B-NVFP4-1M, end to end: 4 of 4 needle-in-a-haystack retrievals at 989k prompt tokens.
- vLLM, tensor parallel over two GPUs
- FP8 KV cache with calibrated static scales
- FlashInfer attention and prefix caching
- MTP speculative decoding
- Native Anthropic Messages API endpoint, used as the backend for Claude Code and opencode
- llama.cpp and GGUF every day on Apple Silicon and CUDA
| Parameter | Test conditions | Typical |
|---|---|---|
| Top-1 agreement vs BF16 | INT8 AutoRound W8A16, teacher-forced | 97.3 % |
| Needle retrieval | NVFP4-1M, 989k prompt tokens, 2× RTX 5090 | 4/4 hits |
| Calibration mix | Chat, prose, code and math | 512 × 4,096 tokens |
| Evaluation stream | Top-24 KL divergence and ΔNLL vs BF16 and FP8 references | ~1M tokens |
| Agentic benchmark | DeepSWE, long-horizon tasks in Docker sandboxes | 113 tasks |
Apps
KalemPRODUCTION
A Qur’an app that answers questions only from the Qur’an and its commentary, and shows the verses behind every answer. Prayer times, Qibla, hadith library, memorization. English, Turkish, and French.
SwiftUI · StoreKit 2 subscriptions · on-device tafsir · Apple Search Ads
Sivas TractorPRODUCTION
A marketplace for new and used tractors and farm equipment. A native SwiftUI rewrite of the original Flutter app, with a Kotlin/Compose Android port and a web storefront.
SwiftUI · Kotlin/Compose · Firebase Auth + Firestore rules · Cloudinary
ReelmIN DEVELOPMENT
An AI video-generation app with a credit economy.
SwiftUI · Swift 6 concurrency · Firebase · StoreKit 2 credits · Fal AI
Skills
| No. | Name | Type | Description |
|---|---|---|---|
| 1 | PY | LANG | Python: quantization, calibration, and evaluation pipelines |
| 2 | SWIFT | LANG | Swift and SwiftUI: iOS and macOS apps, Swift 6 concurrency |
| 3 | KT | LANG | Kotlin and Jetpack Compose: Android apps |
| 4 | TS | LANG | TypeScript and JavaScript: web, Electron, Workers |
| 5 | SQL | LANG | SQL: SQLite, Postgres |
| 6 | TORCH | ML | PyTorch and Transformers |
| 7 | QUANT | ML | llm-compressor, NVIDIA ModelOpt, AutoRound: NVFP4, FP8, INT8, GPTQ, AWQ |
| 8 | EVAL | ML | KL divergence and ΔNLL evaluation, calibration-dataset design |
| 9 | VLLM | SERVE | vLLM multi-GPU serving, FP8 KV cache, speculative decoding |
| 10 | GGUF | SERVE | llama.cpp and GGUF on Apple Silicon and CUDA |
| 11 | CUDA | SERVE | Multi-GPU CUDA inference on Linux |
| 12 | DOCKER | OPS | Docker, Linux server administration, Git, shell |
| 13 | BAAS | PLATFORM | Firebase and Supabase |
| 14 | CF | PLATFORM | Cloudflare Workers and Pages |
| 15 | STORE | PLATFORM | StoreKit 2 and App Store Connect operations |
| 16 | 3D | TOOL | Blender, 3ds Max, ZBrush, AutoCAD, Substance Painter |
About
I studied Industrial Engineering at Istanbul Technical University. Before that, I studied English in London and in Toronto. Industrial engineering still shapes how I work: I measure first, find the bottleneck, and ship on schedule.
I taught myself software by building things I needed. Today my work has two tracks. The first is making large open-weight models run well on local GPUs. The second is shipping native apps to the App Store and building software for small businesses.
- EducationBSc Industrial Engineering, Istanbul Technical University, 2019–2024
- Language studyEmbassy English: Toronto, 2018 · London, 2012
- 2023–2025Industrial engineer, Erdogan Architecture: interior projects for clients in Switzerland and France, 3D visualization, site management
- InternshipsDeha Advertisement (marketing) · Ferre / Femas Metal (production)
- Outside work3D art · DJ and music production
- LanguagesTurkish (native) · English (C2) · German (A1)
| Date | Change |
|---|---|
| 2026-08-23 | Qwen3.8-27B INT8 AutoRound (community variant) released on Hugging Face (INT8 W8A16, group 128). |
| 2026-08-20 | Qwen3.8-27B-NVFP4-1M released on Hugging Face (NVFP4 + FP8, FP8 KV cache). |
| 2026-06-04 | gemma-4-12B-it-NVFP4 released on Hugging Face (NVFP4). |
| 2026-05-02 | Qwen3.6-27B-NVFP4 released on Hugging Face (NVFP4 W4A4). |
| 2026-03-26 | Sivas Tractor released on the App Store. |
| 2026-03-10 | Qwen3.5-27B-NVFP4 released on Hugging Face (NVFP4). |
| 2026-01-31 | Kalem released on the App Store. |
| 2024 | BSc Industrial Engineering, Istanbul Technical University. |
Client work
| Design | Description |
|---|---|
| Accounting software for macOS | SwiftUI and SwiftData, a 261-test suite, and an Excel reporting engine. In daily use at a 5-employee firm. |
| Offline cash-book | Electron desktop app with an encrypted SQLite store. In daily use at the same firm. |
| Websites and web storefronts | TypeScript sites on Cloudflare Workers and Pages, with Firebase or Supabase back ends. |
Contact
| Part | Description |
|---|---|
| BE-2026-APP | Custom iOS or Android app, from design to App Store release |
| BE-2026-LLM | Model quantization, evaluation, or private local inference setup |
| BE-2026-WEB | Website or business software |
| BE-2026-HIRE | Full-time or contract role |
Write to [email protected], or use the form.