ManniX-ITA/Qwen3.8-27B-Omnimerge-v6

Model Overview

ManniX-ITA/Qwen3.8-27B-Omnimerge-v6 is a 27 billion parameter model built upon the Qwen3.8-27B base. It is a task-arithmetic merge of three Qwen3.6 fine-tunes, specifically designed to leverage the advancements of the newer Qwen3.8 generation while incorporating specialized capabilities from its source models. The merge process uses an omnimerge_v2 method, applying deltas from Qwen3.6-based fine-tunes to the Qwen3.8 base.

Key Capabilities & Performance

• Enhanced Code Generation: The model shows a notable improvement in code generation, with LiveCodeBench scores increasing by +6.5 percentage points (0.883 vs 0.818) compared to its v4 predecessor.
• Superior Tool-Calling Safety: In tool-eval-bench hardmode, v6 achieves 156.4 total points, outperforming the Qwen3.8 base and v4. Crucially, it exhibits significantly fewer safety-critical failures, particularly in scenarios like Cross-Turn Sleeper Injection, which other models in its cohort consistently fail.
• Reasoning & General Capability: The model integrates reasoning capabilities from its source fine-tunes, including those distilled from Claude-Opus and specialized reasoning anchors.
• Vision Tower & MTP Head: It retains the stock Qwen3.8 vision projector and deliberately preserves the 3.8 MTP head for optimal decode speed, avoiding the integration of 3.6-trained MTP deltas.

When to Use This Model

This model is particularly well-suited for applications requiring:

• Robust code generation and understanding.
• Reliable and safe tool-calling or function-calling capabilities.
• General-purpose reasoning tasks, benefiting from its merged fine-tunes.

It offers a strong balance of performance and safety, making it a compelling choice for developers looking for an advanced Qwen-based model with specialized enhancements.