groxaxo/Code-Writer-V2-Obliterated-BF16

Overview

Code Writer V2 — Obliterated (BF16) is a 27-billion-parameter model by groxaxo, representing the full-fidelity, BF16 precision version of the Code Writer V2 series. It is a merged model, combining llmfan46/Qwen3.5-27B-Writer-V2-uncensored-heretic with a purpose-trained coding_mix_8k LoRA. This model is designed to excel at both creative writing and code generation, offering a unique blend of capabilities.

Key Capabilities

• Hybrid Architecture: Built on Qwen3.5, it features 64 decoder layers, with 16 carrying full attention and the rest utilizing GDN linear attention for extended memory.
• Long Context: Supports an impressive context length of up to 200,000 tokens, enabling it to handle entire codebases or manuscripts.
• Vision-Capable: Includes a full vision tower, though vision capabilities are noted as present but untested.
• Uncensored & Free: Described as "heretic by lineage and free by intent," it operates without flinching or lecturing.
• Full Precision: This BF16 version is the reference master, with no quantization, making it suitable for maximum quality or as a base for further quantization.

Which Version to Use?

• This BF16 version: Ideal for golden reference, further quantization, or when maximum quality is paramount.
• FP8 version: Recommended for day-to-day serving on vLLM, offering a smaller footprint (~half the size) while maintaining high fidelity.

Good For

• Writing: Fiction, screenplays, copy, and long-form prose.
• Code Generation: The integrated coding LoRA is specifically trained for this purpose.
• Long-form Work: Its 200k token context window makes it suitable for extensive documents, codebases, or conversations.