aisquared/bolt-instruct-v2
aisquared/bolt-instruct-v2 is a 30 billion parameter multimodal causal transformer model developed by aisquared, fine-tuned from meta-models/Muse-Glimmer-30B with a 131,072+ token context length. It is specifically optimized for concise reasoning, accurate tool selection, and efficient execution of multi-turn agentic workflows, supporting both text and image modalities. This model excels at instruction following, structured output, and code generation, aiming to reduce cost and latency compared to other 30B models in agentic tasks.
Loading preview...
Bolt Instruct V2: Optimized for Agentic Workflows
Bolt Instruct V2 is a 30 billion parameter multimodal causal transformer model developed by aisquared, fine-tuned from meta-models/Muse-Glimmer-30B. It is specifically designed for concise reasoning, accurate tool selection, and efficient execution of tool-based workflows across multi-turn agentic scenarios, supporting both text and image modalities with a context length of 131,072+ tokens.
Key Capabilities
- Strong Instruction Following: Excels in adhering to instructions, particularly in multi-turn dialogues.
- Efficient Agentic Workflows: Optimized for automated task completion and agentic coding, reducing unnecessary deliberation.
- Concise Reasoning: Demonstrates significantly fewer reasoning tokens and lower latency in agentic benchmarks compared to its base model and Gemma 4 31B.
- Multimodal Support: Handles both text and image inputs.
- Structured Output: Capable of generating structured data like JSON.
- Code Generation: Supports code assistance and generation.
Good For
- Chat Assistants & Enterprise Copilots: Ideal for building interactive AI agents that require precise instruction adherence and tool use.
- RAG Pipelines & Document Processing: Effective for summarization, extraction, and transformation tasks.
- Automated Task Completion: Suited for scenarios requiring multi-step agentic execution with tool calls.
- Cost and Latency Reduction: Offers improved efficiency in agentic tasks compared to other models in its size class.
Evaluation shows Bolt Instruct V2 significantly reduces reasoning and output tokens, and improves latency on benchmarks like BFCL and τ²-bench, while also achieving a 79% success rate on ACEBench parallel function calling.