DeepSeek-Coder-V2-Lite-Instruct-FLOW 🌊

High-Performance Streaming Model Container (.gguf.flow) for FLOW Studio
Runs on consumer 16GB Apple Silicon Macs and Windows PCs under a fixed 2.5 GB RAM footprint with Zero OOM.

---

⚠️ Important Compatibility Notice

This model is packaged in the FLOW-Boost container format (.gguf.flow), optimized for out-of-core streaming and in-flight decompression.

• Compatible Runtime: FLOW Studio (macOS & Windows)
• Standard GGUF Version: If you are using standard llama.cpp, Ollama, or LM Studio, please download the standard GGUF build from:
  👉 masahiroid/DeepSeek-Coder-V2-Lite-Instruct-Pyrite-2bit-GGUF

---

Model Features

1. RAM-NVMe Decoupled Streaming: Decouples RAM from SSD storage. Keeps active memory strictly bounded to ~2.5 GiB, completely eliminating Out-Of-Memory (OOM) crashes on 16GB Macs.
2. In-Flight Byte-Shuffle + Zstandard Codec: Implements in-flight byte-plane rearrangement and Zstd decompression targeting Apple Silicon unified memory pipelines. (Note: On 2-bit quantized weights, entropy limits size compression to ~8-12%, whereas unquantized FP16/BF16 weights achieve ~31.5% reduction).
3. Prior Art & Engineering Scope: Builds upon principles from Apple LLM in a Flash (Alizadeh et al., 2023) and byte-shuffling lossless filters (Blosc, IBM ZipNN), specialized for Metal unified memory slicing.
4. Lossless Quantized Baseline: Preserves exact 2-bit Pyrite sensitivity-guided precision ($\epsilon \equiv 0.0$ bitwise exact match).

---

How to Run

1. Download FLOW Studio

Download the signed macOS installer or Windows launcher from GitHub: 👉 FLOW Studio Releases

2. Launch & Select Model

Place DeepSeek-Coder-V2-Lite-Instruct.gguf.flow into your models_flow/ directory, open FLOW Studio, and start local inference with full OpenAI-compatible API support (http://localhost:8080/v1).

---

Attribution

• Developer: Masahiro NAKATSUGAWA (中津川 昌弘) / ORCID: 0009-0000-3224-0103
• Organization: AntenneFrance / n.p.minerva & Co.