Qwevolution-V0-27B

Qwevolution V0 27B

<!-- polished-overview:start -->

Overview

Qwevolution-V0-27B is a model checkpoint packaged for compatible Hugging Face runtimes, published by groxaxo. It is intended for open-source evaluation, reproducible experimentation, and compatible local or hosted inference workflows. The wording below is deliberately limited to what can be verified from this repository's metadata and artifacts.

At a glance

What is included

• *.safetensors (12 files)
• config.json
• generation_config.json
• tokenizer.json
• tokenizer_config.json
• processor_config.json
• chat_template.jinja
• Additional configuration, tokenizer, processor, or shard files (20 visible artifacts total)

Quick start

Getting started

Start with the upstream library named in the repository metadata and keep all configuration, tokenizer, processor, and weight files together. This repository is an artifact release, so the source project remains the authoritative reference for task-specific loading code.

Compatibility and responsible use

• Use a runtime that explicitly supports this format, architecture, and modality.
• Keep configuration, tokenizer, processor, projection, and weight files from the same revision together.
• Review the source model card and license before redistribution or deployment.
• Hardware needs depend on parameter count, context length, cache precision, quantization, and concurrency.
• Report reproducible issues with the runtime version, hardware, launch command, and a minimal example.

Generated outputs may be inaccurate or unsuitable for a given use case. Users are responsible for testing behavior, applying appropriate safeguards, and complying with applicable licenses and laws.

<!-- polished-overview:end -->

A weight-space blend of the ConicCat/Qwen3.5-27B-Writer-V2 vision-language model with the groxaxo/Qwen3.5-27B-Writer-V2-Heretic-Coding-LoRA coding LoRA adapter, merged with peft.merge_and_unload(safe_merge=True) in bfloat16.

Composition

• Base model: ConicCat/Qwen3.5-27B-Writer-V2 — Qwen3_5ForConditionalGeneration, 27B, BF16, license apache-2.0. Base commit 55f457bed78b68442d25815bd156b98f2e6fd341.
• LoRA adapter: groxaxo/Qwen3.5-27B-Writer-V2-Heretic-Coding-LoRA — r=16, α=32. Adapter commit 0b3508764f80741f612c251c7f47b1e210ac2e63.

Deliberate cross-base merge. The coding LoRA was originally trained on llmfan46/Qwen3.5-27B-Writer-V2-uncensored-heretic. It was intentionally merged onto the clean ConicCat/Qwen3.5-27B-Writer-V2 (same Writer-V2 lineage / identical architecture, vocab and module tree).

What was merged (and what wasn't)

The adapter touched 607 modules. Of these, 496 language-tower modules carried trained (non-zero) weights and were merged into the base. The remaining 111 vision-tower modules were zero in the adapter itself (never trained) — merging them is a no-op, so the vision tower is identical to the base model and vision behaviour should match ConicCat/Qwen3.5-27B-Writer-V2.

Provenance & verification

• Merged on CPU in bfloat16 with safe_merge=True (per-layer NaN check).
• Full adapter coverage verified before merge; trained language weights confirmed loaded (non-zero B).
• Integrity verified by reloading the saved checkpoint and running a forward pass (finite logits).
• See merge_manifest.json for exact base/adapter commits and library versions.

Usage

from transformers import AutoModelForImageTextToText, AutoProcessor
model = AutoModelForImageTextToText.from_pretrained("groxaxo/Qwevolution-V0-27B", dtype="bfloat16", device_map="auto")
processor = AutoProcessor.from_pretrained("groxaxo/Qwevolution-V0-27B")