Bolt Instruct V2: Flagship Multimodal Model for Optimized Agentic Execution

Introduction

At AISquared we build agentic systems for enterprise workflows, and running them at scale makes one thing quickly apparent: the dominant cost is rarely a lack of capability. Modern models are generally good enough at deciding what to do. Where they get expensive is in the deliberation while performing a task.

Bolt Instruct V2 is AISquared’s flagship multimodal model designed for agentic AI applications where successful execution depends not only on generating accurate responses but also on reasoning through tasks, selecting appropriate tools, and coordinating actions across multiple steps. Developed as a supervised fine-tune of Muse Glimmer 30B, Bolt Instruct V2 builds on the underlying model’s text, image, and video capabilities while placing greater emphasis on efficient reasoning and reliable tool use. The objective is to provide organizations with a model capable of supporting complex automated workflows while reducing the computational overhead associated with repeated reasoning and tool execution.

A central focus of Bolt Instruct V2 is improving how models reason about and execute tool-based tasks. Agentic workflows frequently require models to determine which tools are needed, construct appropriate tool calls, interpret intermediate results, and continue executing until a task is complete. These interactions can involve multiple independent operations, making the ability to issue parallel tool calls particularly valuable. Bolt Instruct V2 was fine-tuned to produce concise reasoning before taking action, select tools accurately, and support multiple tool calls within a single turn. By reducing unnecessary deliberation and enabling concurrent execution of independent operations, the model is designed to make multi-step workflows more efficient without sacrificing the capabilities required to handle complex tasks.

Our evaluations demonstrate improvements in several of these areas relative to the underlying Muse Glimmer model. Bolt Instruct V2 achieved stronger performance across parallel tool-calling benchmarks and instruction-following evaluations, while using 49.8% fewer reasoning tokens and achieving 36.2% lower mean latency on measured workloads. These improvements were accompanied by a 36.4% reduction in benchmark processing costs, highlighting the potential benefits of optimizing models specifically for agentic execution. Performance was not uniformly higher across every benchmark, however, and the evaluations also identified areas where additional improvements remain necessary.

Designing Bolt Instruct V2 for fewer Reasoning tokens, lower cost and lower latency

The design of Bolt Instruct V2 started with a practical question: how could we reduce the cost of running an agent without losing the tool-use capabilities it needed? That question guided our choice of base model, the data we assembled, the training method, and ultimately how we evaluated the result.

We evaluated four candidates before choosing a base: Gemma 4 31B, Laguna-XS-2.1, Llama-3.1-Nemotron-Nano-8B, and Muse Glimmer 30B. We needed capable tool use, practical serving costs, and a separate reasoning channel that would let us target deliberation directly during fine-tuning.

Nemotron Nano fell short on capability. The 8B model scored 31.5% on airline, 23.1% on retail, and 0% on banking in our τ²-bench evaluations, placing it well below the 30B-class candidates.. Laguna-XS-2.1 was stronger, but at 33.4B parameters it was larger than the other candidates and text-only, leaving out the multimodal support our workloads require.

Gemma 4 31B was the closest contender and remains our external reference. Its multi-turn accuracy was comparable to Muse Glimmer’s, but it generated roughly 94% more reasoning tokens per task across the benchmarks we measured. Muse Glimmer therefore offered a more efficient starting point for the behavior we wanted to train.

Muse Glimmer 30B also fit our deployment and training setup. Its 29.8B parameters occupy just under 60 GB in bf16, supporting room for the KV cache. Its vision encoder is separate from the 52-layer text decoder, allowing us to freeze the visual component while adapting the language layers. Crucially, its dedicated reasoning channel and graded reasoning_strength parameter gave us direct access to the behavior we wanted to change.

Training Data

The  selected corpus demonstrates efficient tool use across complete workflows. Each session required tool schemas, at least one reasoning block, and a genuine user request. Reasoning remained brief but consistently present, with longer deliberation included where needed. We also explicitly included parallel tool calls to address Muse Glimmer’s 0.00% baseline scores on BFCL’s parallel categories.

Optimized with Response-only masking and Lora

We fine-tuned Muse Glimmer 30B with LoRA to encourage brief reasoning followed by precise tool calls, grouping independent calls for parallel execution.We moved to supervised fine-tuning after DPO increased reasoning-token usage while reducing accuracy. We wanted brief but nonzero reasoning, so two-thirds of our preference pairs favored shorter reasoning and one-third favored longer reasoning. These competing signals appeared to shift behavior without achieving the intended length reduction, while single-turn preferences provided limited supervision for completing long episodes. SFT lets us demonstrate the desired reasoning lengths and execution behavior directly across complete sessions. The corpus contained 7,646 multi-step agentic sessions, each including tool schemas, reasoning, and a genuine user request, giving the model examples of efficient tool use across complete workflows.

We applied loss only to assistant reasoning, tool calls, and final answers using Unsloth’s train_on_responses_onlywith the Muse chat template’s channel markers. User messages, tool schemas, and tool results remained context, keeping supervision focused on the outputs the model needed to generate. LoRA adapters targeted the language layers’ attention and MLP projections with rank 32, alpha 32, and zero dropout. We kept the vision tower frozen because the corpus contained no visual data.

Training ran for one epoch on two H200 GPUs, with batch size 1 and four gradient accumulation steps. We used adamw_8bit at a learning rate of 2 × 10⁻⁴, with a linear schedule, five warmup steps, and weight decay of 0.001. A maximum sequence length of 65,536 tokens accommodated long interaction histories so the model could learn from sustained multi-step execution.

Model Characteristics

Efficient Reasoning and Lower Operating Costs

Agentic workflows often require repeated cycles of reasoning, tool selection, execution, and interpretation before a task can be completed. While these intermediate steps are essential for coordinating complex operations, generating excessive reasoning tokens can introduce additional computational overhead, particularly in applications that require frequent model inference. Bolt Instruct V2 was fine-tuned to emphasize concise, task-directed reasoning, allowing the model to focus on identifying the appropriate next action and constructing the necessary tool calls without unnecessary deliberation. Across our measured workloads, this approach resulted in 49.8% fewer reasoning tokens and a 36.4% reduction in benchmark processing costs relative to Muse Glimmer 30B. These improvements are particularly relevant for enterprise applications that execute large volumes of agentic workflows, where reducing the computational requirements of individual inference steps can contribute to more economical operations.

Faster Responses and Parallel Tool Execution

Many enterprise workflows involve multiple independent operations, such as retrieving customer information, checking inventory availability, and evaluating delivery options. When these operations are executed sequentially, the overall workflow can become constrained by the cumulative latency of individual tool interactions. Bolt Instruct V2 supports up to six independent tool calls within a single turn, allowing applications to initiate multiple operations concurrently when their dependencies permit. Combined with its emphasis on concise reasoning, this capability supports more efficient coordination of tool-based tasks and reduces unnecessary sequential execution. In our evaluations, Bolt Instruct V2 achieved 36.2% lower mean latency and 50.9% lower 95th-percentile latency compared with Muse Glimmer 30B on measured workloads. These results demonstrate improvements in both typical response times and the latency of slower requests, which can be particularly important for interactive applications and workflows involving repeated model interactions.

Multimodal Capabilities and Long-Context Support

Enterprise AI applications increasingly require models to interpret information across multiple modalities while maintaining context throughout extended workflows. Bolt Instruct V2 builds on the multimodal capabilities of Muse Glimmer 30B, supporting text, image, and video inputs within a unified model architecture. This allows applications to incorporate information from sources such as documents, screenshots, visual records, and video content without requiring separate models for each input modality. The model also supports a 131K-token context window, providing capacity for lengthy documents, detailed tool schemas, extensive interaction histories, and intermediate results generated during multi-step execution. Together, these capabilities enable Bolt Instruct V2 to support agentic workflows that require both multimodal understanding and the ability to maintain substantial contextual information across successive reasoning and tool-execution steps.

Evaluation

We wanted to understand how Bolt Instruct V2’s shorter reasoning affected both efficiency and task performance. We compared it with Muse Glimmer 30B and Gemm 4 31B across single-turn tool calling, multi-turn agentic execution, and general capability, while also measuring inference efficiency. Each benchmark ran through its official harness, using Muse Glimmer’s published sampling settings,temperature 1.0, top_p 0.95, and top_k 64,applied server-side through vLLM.

Efficiency

We measured efficiency across BFCL, τ²-bench, and AutomationBench to capture single-turn tool calls, multi-turn dialogue, and long-horizon enterprise workflows. Across these benchmarks, Bolt Instruct V2 used 42.2%–68.2% fewer reasoning tokens and 23%–40% fewer output tokens than Muse Glimmer 30B. The largest reasoning reduction came on AutomationBench, where usage fell from 6,351 to 2,022 tokens per task.

Gemma 4 31B achieved higher multi-turn scores but used substantially more reasoning: 2,609 tokens versus Bolt’s 635 on BFCL, and 10,835 versus 2,022 on AutomationBench. Latency showed a more mixed picture. Bolt was faster than its base on both measured benchmarks, reaching 12.60 seconds on BFCL and 51.4 seconds on τ²-bench. Gemma was fastest on BFCL at 7.47 seconds, but took 129.8 seconds on τ²-bench, highlighting the importance of comparing efficiency alongside task performance.

Figure 1. Reasoning token, latency, and cost reductions with Bolt Instruct V2.

Agentic — multi-turn

On τ²-bench, Bolt Instruct V2 uses roughly half as many steps as its base model while scoring 76.83 at temperature 0. That makes execution shorter and more efficient. For these longer conversations, the trade-off is fewer steps for an 8.54-point drop in task success.

Agentic — single-turn tool calling

Single-turn tool calling shows the clearest effect of the fine-tune, with the largest gains where the base model was weakest. BFCL Non-Live shows a 2.4× improvement on structured multi-call tasks. ACEBench parallel function calling shows a 31-point gain that brings Bolt Instruct V2 within two points of Gemma 4 31B at 81%, despite Gemma spending roughly 4 times the reasoning tokens to get there.

BFCL Live, which tests tool selection against real-world API schemas, improves from 55.84% to 59.67%. The pattern across all three is consistent: training the model to reason concisely and group independent calls produced better tool selection and formatting as well, not only faster execution.

General capability

General capability held up under the fine-tune, and in two areas improved markedly: BBH rose from 63.00% to 89.00% and GPQA Diamond from 33.6% to 57.5%, with IFEval prompt-strict at 73.75% against the base model’s 65.06%. ARC Challenge and TruthfulQA MC2 stayed within two points. Training for efficiency often costs general capability; here it did not.

Bolt Instruct V1 and V2 are not successive versions of the same model. V1 is built on Olmo 3 (32B, 64K context, text-only) and optimized for instruction following; V2 is built on Muse Glimmer (30B, 128K context, multimodal) and optimized for agentic tool execution. Each should be read against its own foundation model, but Bolt Instruct V2 improves on Muse Glimmer’s IFEval and BBH.

Conclusion

Bolt Instruct V2 shows that a focused training objective can make agentic execution substantially more efficient. A single epoch over 7,646 sessions with rank-32 LoRA reduced reasoning-token usage by 42–68%, alongside gains in instruction following and parallel tool calling. These results demonstrate the value of directly teaching concise reasoning and tool use through supervised fine-tuning.

The benefit is clearest on bounded, tool-driven tasks, where Bolt delivers shorter execution and lower per-task cost. The base model retains an advantage on longer workflows in τ²-bench and AutomationBench, so efficiency gains need to be considered alongside task completion.