Azure99/Blossom-V7-9B

Blossom-V7-9B: A Multimodal Model for Agentic Workflows

Blossom-V7-9B is part of the Blossom-V7 family of open-weight, general-purpose multimodal models developed by Azure99, designed for efficient local deployment. Built upon the Qwen3.5-9B base, this model integrates advanced capabilities for diverse applications, from general conversation to complex agentic tasks.

Key Capabilities

• Efficient Adaptive Thinking: Employs an always-on adaptive thinking mechanism that dynamically adjusts reasoning depth based on task difficulty, producing high-quality results with significantly shorter reasoning traces compared to Qwen3.5 and Qwen3.6.
• Tool Use with Interleaved Thinking: Enhances agentic performance by reasoning and making decisions before each tool call, facilitating robust interaction with external tools.
• Image Understanding: Processes and understands image inputs alongside textual information, enabling multimodal interactions.
• Extended Context Length: Supports a maximum context window of 262,144 tokens, with 131,072 tokens recommended for optimal performance.
• Faster Inference: Incorporates Multi-Token Prediction (MTP) for speculative decoding, supported in both vLLM and llama.cpp, to accelerate inference.
• Broad Application Scope: Post-trained for general assistant use, covering everyday conversation, world knowledge, mathematics, reasoning, coding, web development, and data visualization.

Usage Considerations

Blossom-V7-9B utilizes a custom chat template; users must employ the bundled chat_template and avoid substituting or combining it with Qwen3.5 templates. For multi-turn conversations, it is crucial to replay the assistant's complete reasoning along with its answer to maintain performance in subsequent turns. The model is optimized for memory-constrained GPUs, mobile devices, and lighter workloads, offering a resource-efficient option within the Blossom-V7 series.