Executive Summary
AISquared is introducing Bolt Instruct V2 and Bolt CLS, two additions to the Bolt model family designed to make enterprise agentic AI more capable, efficient, and secure. Bolt Instruct V2 improves agentic execution through more efficient reasoning and parallel tool calling, achieving 42.2% – 68.2% fewer reasoning tokens, 23.2% – 39.6% fewer output tokens, 36.2% lower mean latency, and 36.4% lower cost per task completion compared with its base model, Meta’s Muse Glimmer. Complementing these capabilities, Bolt CLS introduces lightweight models for safety detection, sensitive information identification, and intelligent model routing, demonstrating 99.90% recall in PII detection and 94.19% effective routing accuracy in our evaluations. Together, these releases advance AISquared’s vision for modular AI architectures that move beyond monolithic agents, combining specialized capabilities to deliver more efficient, flexible, and controlled enterprise AI systems.
Enterprise AI Has a Systems Problem
The enterprise AI conversation has largely focused on building more powerful models: models that can reason more deeply, execute more complex tasks, and interact with an expanding ecosystem of tools and applications. As organizations move from AI assistants to autonomous agents, a different challenge is emerging: how do we make these increasingly capable systems efficient, secure, and practical to operate?
Today, many AI applications rely on what we call monolithic agents: highly capable models given broad responsibility for interpreting requests, selecting tools, and executing workflows. While these architectures can be effective, they often introduce unnecessary complexity. Simple tasks are routed through expensive models, specialized decisions are handled through generative inference, and agents may receive access to far more capabilities than individual workflows require.
At AISquared, we believe the answer isn’t simply to build larger or more capable models. It’s to build more intelligent systems.
That means moving beyond monolithic architectures toward modular systems that combine specialized models, each responsible for what it does best. Effective agentic AI requires models that can reason, interact with tools, and execute complex workflows, alongside supporting infrastructure that can evaluate incoming requests, identify potential risks, and select appropriate downstream capabilities based on operational and business requirements.
Our latest Bolt releases address both sides of this challenge.
Bolt Instruct V2 is our flagship multimodal model, optimized for agentic execution. Building on Muse Glimmer, it emphasizes efficient reasoning, reliable tool selection, and parallel tool execution to improve how AI models perform complex, multi-step tasks in fewer steps and at lower cost. Its relatively compact size also supports deployment across a range of hardware environments, including appropriately provisioned edge infrastructure.
Bolt CLS is a new subfamily of lightweight classification models designed to support AI safety, sensitive information detection, and intelligent model routing. Rather than relying on generative inference for these specialized decisions, Bolt CLS uses purpose-built classifiers to evaluate incoming requests and produce structured signals that applications can act upon.
While these models serve different purposes, they share a common objective: making enterprise AI more effective by matching the right capabilities to the right responsibilities. Bolt Instruct V2 focuses on executing work efficiently. Bolt CLS helps applications determine how that work should be evaluated and where it should be directed. Together, they represent our approach to building a more modular, intelligent, and practical foundation for enterprise agentic AI.
Bolt Instruct V2: More Capability, Less Overhead
As enterprises continue to adopt agentic AI, model performance can no longer be measured solely by the quality of generated responses. Successful agents must interact with tools, make decisions across multiple steps, and execute tasks efficiently. A model that produces accurate answers but requires excessive reasoning, unnecessary tool interactions, or long execution times can create significant operational challenges when deployed at scale. Bolt Instruct V2 was developed to address these challenges.
Built through targeted supervised fine-tuning of Muse Glimmer, Bolt Instruct V2 is optimized for concise reasoning, accurate tool selection, and efficient execution of tool-based workflows. A central feature of the model is its ability to issue multiple independent tool calls within a single turn. Rather than processing every operation sequentially, Bolt Instruct V2 supports multiple tool calls in parallel within a single step , enabling applications to initiate independent operations concurrently when appropriate. We chose Muse Glimmer as the base model because, at the time of this writing, it is the most capable model of U.S. origin at agentic execution, instruction following, and visual reasoning at or below the 30-billion parameter size class, making it easy to deploy and use on-device and in resource-constrained on-prem environments.
Consider an enterprise application tasked with gathering customer information, checking inventory availability, and evaluating delivery options. These operations may involve separate systems but do not necessarily need to happen one after another. By coordinating independent tool calls, an agent can reduce unnecessary sequential processing and execute workflows more efficiently.
Our evaluations demonstrate meaningful improvements in both capability and inference efficiency. Compared with the underlying Muse Glimmer 30B model, Bolt Instruct V2 achieved:
- 42.2%–68.2% fewer reasoning tokens, reducing the computational overhead associated with intermediate reasoning.
- 36.2% lower mean latency, improving response times across evaluated workloads.
- 36.4% lower cost per task, demonstrating the economic benefits of more efficient inference.
- 23.2%–39.6% fewer output tokens, minimizing generated text generation needed during task completion
These efficiency gains were accompanied by improvements in tool-calling performance. On ACEBench, Bolt Instruct V2 improved from 48% to 79%,while the BFCL parallel tool-calling evaluation demonstrated substantial gains, rising from 10.1% to 24.4% over the base model. Broader reasoning improved notably in BBH from 63.0% to 89.0%, while general knowledge held steady, matching Bolt Instruct V1 on ARC Challenge.
While there is some degradation versus the base model on completion rates for the longest muIti-turn agentic workflows, the benefits to conciseness trained into Bolt Instruct V2 provide a clear improvement for the vast majority of real-world uses for the model in the AISquared UNIFI platform.
AI agents frequently perform repeated cycles of reasoning, tool selection, and execution. Even modest reductions in the computational requirements of individual steps are meaningful when applications operate at scale. For organizations operating AI agents across large numbers of workflows, greater inference efficiency can translate into faster execution, more predictable performance, and better control over operating costs.
Bolt CLS: Intelligence Before Execution
While Bolt Instruct V2 focuses on executing complex workflows, effective enterprise AI systems must also make decisions before execution begins. Consider the range of questions an AI application may need to answer when receiving a request:
- Does the request contain potentially unsafe content?
- Is someone attempting to bypass the system’s safeguards?
- Does the request include personally identifiable information?
- Which available model or agent is best equipped to handle the task?
These are important decisions, but they do not necessarily require a large generative model to answer them. In many cases, they are fundamentally classification problems, not content generation problems. That’s the simple idea behind Bolt CLS, our new family of lightweight models designed for AI safety and intelligent routing.
Unlike generative language models that produce textual responses, Bolt CLS models use specialized classification architectures to evaluate requests and return structured predictions. This allows applications to incorporate safety, privacy, and routing decisions into their workflows without requiring autoregressive text generation for each classification task. Furthermore, framing these problems as predictive rather than generative results in much smaller, more efficient models. Whereas generative models typically range in the billions of parameters, our Bolt CLS models range from approximately 100 to 400 million parameters. This reduction in size allows users to utilize these models across a much wider range of hardware options, including CPU-only workloads.
The Bolt CLS family includes four specialized models:
| Model | Purpose |
|---|---|
| Bolt CLS Content | Identifies requests containing potentially unsafe or policy-violating content. |
| Bolt CLS Jailbreak | Detects attempts to circumvent model instructions or safety mechanisms. |
| Bolt CLS PII | Identifies requests containing personally identifiable information. |
| Bolt CLS Router | Directs requests toward appropriate downstream models based on task requirements and available capabilities. |
Together, these models provide a flexible classification layer that applications can use to evaluate incoming requests, incorporate appropriate safeguards, and select downstream capabilities.
Safety and Privacy Screening
As enterprise AI applications interact with sensitive information, internal systems, and external services, organizations need mechanisms for identifying potential risks before requests are processed. Bolt CLS provides specialized detection capabilities that applications can incorporate into their existing safety and privacy workflows.
Our evaluations demonstrate particularly strong results for PII detection. Bolt CLS PII achieved 99.90% recall while maintaining 99.73% precision at a retrospectively examined classification threshold, demonstrating strong detection performance on the evaluated dataset. The Content and Jailbreak models also demonstrated the ability to identify more than 99% of positive examples at high-recall operating thresholds, although achieving that sensitivity introduced higher false-positive rates.
Identifying a potentially problematic request does not necessarily mean that the request should be rejected. Depending on an organization’s policies, a positive classification may instead trigger additional verification, more detailed evaluation, or alternative processing. Bolt CLS supports this flexibility through configurable classification thresholds, allowing applications to balance detection sensitivity against the risk of incorrectly flagging legitimate requests. Rather than prescribing a single approach to safety enforcement, these models provide specialized signals that organizations can incorporate into their own AI governance and operational workflows.
Intelligent Model Routing
One of the most significant challenges facing enterprises deploying AI is determining which models or agents should handle which requests. Different models offer different capabilities, costs, and performance characteristics. A straightforward summarization request may be handled effectively by a smaller model, while a complex analytical task may require deeper reasoning capabilities.
In monolithic-agent architectures, these distinctions are often overlooked. Organizations may direct simple and complex requests through the same powerful models, creating unnecessary inference costs. Agents may also receive broad access to tools and systems regardless of whether those capabilities are required for a particular task. This approach can introduce operational inefficiencies and expand the potential consequences of incorrect or unauthorized tool use.
Bolt CLS Router addresses the model-selection side of this challenge through capability-based routing.
Unlike traditional routing approaches that rely on fixed rules or predefined relationships between tasks and specific models, Bolt CLS Router evaluates requests against the capabilities of available destinations. As organizations introduce new models, update existing infrastructure, or deploy specialized agents, their routing requirements change. Bolt CLS Router allows applications to describe available capabilities dynamically, enabling routing decisions to adapt without requiring entirely new routing logic. Applications can define candidate destinations using information such as their supported tasks, relative inference costs, expected latency, and reasoning capabilities. The Router then selects an appropriate destination based on the request and the application’s configured optimization objective. Organizations can configure routing behavior to prioritize cost efficiency, response latency, output quality, or a balance of these considerations, depending on their operational requirements, utilizing a flexible approach to managing multi-model AI environments.
Across 10,880 held-out routing tasks, Bolt CLS Router achieved 89.81% top-1 routing accuracy. When accounting for cases where multiple destinations were considered capable of appropriately handling the same request, effective routing accuracy increased to 94.19%. This second metric is particularly relevant in enterprise environments, where multiple models may be suitable for a given task.
Put simply, effective routing asks whether the selected destination is considered capable of completing the requested task, even if it wasn’t the single preferred destination. While this metric does not establish that every selection was optimal for cost, latency, or quality, it provides a useful measure of the Router’s ability to identify appropriate processing pathways.
For organizations operating across multiple AI models and agents, capability-based routing offers a foundation for matching requests to available resources while adapting to evolving operational requirements.
Building the Next Generation of Enterprise AI
As AI evolves from conversational assistants to increasingly autonomous systems, the infrastructure supporting these applications must evolve alongside them. Powerful generative models remain essential, but relying on a single model to handle every responsibility is not necessarily the most efficient, flexible, or controlled approach to enterprise AI.
At AISquared, we believe the future lies in intelligent, modular architectures that combine specialized capabilities to address different responsibilities within an agentic workflow. Bolt Instruct V2 and Bolt CLS reflect this philosophy, bringing together three core principles:
- Evaluate intelligently. Use lightweight classifiers to identify potential safety and privacy risks, enabling applications to apply appropriate safeguards before execution.
- Route appropriately. Match incoming requests with downstream models and agents based on their capabilities, task requirements, and operational priorities.
- Execute efficiently. Leverage capable agentic models to perform complex workflows through precise tool selection, concise reasoning, and parallel execution.
Together, these capabilities illustrate how organizations can move beyond monolithic agents toward AI systems that use specialized components to make decisions and execute work more effectively. Bolt CLS provides the classification and routing capabilities that help determine how requests should be handled, while Bolt Instruct V2 delivers the reasoning and tool execution needed to complete complex tasks.
These releases represent our continued investment in making enterprise AI more capable, efficient, and practical to deploy.
Because building better enterprise AI isn’t just about making models more powerful. It’s about building systems that know when, where, and how to use that power.
Want to explore the technical details behind these releases? Read our deep dives into Bolt Instruct V2’s training methodology and evaluation results and Bolt CLS’s approach to AI safety and capability-based routing.
