ToolSandbox
Models use stateful tools in multi-turn conversations with a simulated user. Tools have hidden state dependencies, and runs are scored against milestones.
The repo moved to the apple-aiml-research organization on GitHub.
In the authors’ words
Recent large language models (LLMs) advancements sparked a growing research interest in tool assisted LLMs solving real-world challenges, which calls for comprehensive evaluation of tool-use capabilities. While previous works focused on either evaluating over stateless web services (RESTful API), based on a single turn user prompt, or an off-policy dialog trajectory, ToolSandbox includes stateful tool execution, implicit state dependencies between tools, a built-in user simulator supporting on-policy conversational evaluation and a dynamic evaluation strategy for intermediate and final milestones over an arbitrary trajectory. We show that open source and proprietary models have a significant performance gap, and complex tasks like State Dependency, Canonicalization and Insufficient Information defined in ToolSandbox are challenging even the most capable SOTA LLMs, providing brand-new insights into tool-use LLM capabilities. ToolSandbox evaluation framework is released at https://github.com/apple/ToolSandbox
- Level
- Everyday Most adults, without training
- Scored
- Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
- Answers
- Agent task Many steps in a terminal, browser, desktop, app, game or robot.
LFM2.5-24B-A2B leads with 80.7%
Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.
- 1 LFM2.5-24B-A2B 80.7%
- 2 LFM3-1.7B 80.4%
- 3 LFM2.5-2.6B 77.8%
- 4
Nemotron 3.5 Bolt E4B 74.1% - 5 LFM2.5-VL-3B 59.5%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | LFM2.5-24B-A2B | 80.7% | Oct 2026 | Model card | |
| 2 | LFM3-1.7B | 80.4% | Oct 2026 | Model card | |
| 3 | LFM2.5-2.6B | 77.8% | Aug 2026 | Model card | |
| 4 | Nemotron 3.5 Bolt E4B | 74.1% | Oct 2026 | Model card | |
| 5 | LFM2.5-VL-3B | 59.5% | Aug 2026 | Model card |
2 models have held number one since its release in Aug 2024
Lab claims count from the day the card came out, independent runs from the model’s release.
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
2 labs put it in a launch post in the last six months
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| 27 Oct 2026 | LFM3-1.7B | 80.4 | ToolSandbox |
| 23 Oct 2026 | Nemotron 3.5 Bolt E4B | 74.1 | ToolSandbox |
| 13 Oct 2026 | LFM2.5-24B-A2B | 80.7 | ToolSandbox |
| 12 Aug 2026 | LFM2.5-VL-3B | 59.5 | ToolSandbox |
| 4 Aug 2026 | LFM2.5-2.6B | 77.8 | ToolSandbox |
One system card reports it
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| 23 Oct 2026 | Nemotron 3.5 Bolt E4B technical report | 74.1 | ToolSandbox, released Q4_K_M imatrix checkpoint (Table 5). BF16 68.77 (Table 3) | Capabilities |
More in Tool use
- AutomationBench In use Agent completes end-to-end business workflows inside a simulated company exposing dozens of REST APIs. Graded on resulting state.
Gemini 4 Carbon 53.9% - MCP Atlas In use Multi-step workflows over real MCP servers. Scores tool-use competency on realistic tasks.
Muse Spark 1.1 88.1% - BFCL V4 In use Models call functions and tools, including multi-turn agentic tasks with web search and memory. Checked against expected calls and outcomes.
Qwen3.7-Max 75% - τ³-bench Banking In use Agents act as banking customer-service reps over many turns with a simulated user, following policy. Pass^1 on the final database state. Qwen 3.8 Max 55.1%
- Claw-Eval In use 300 human-verified agent tasks (service orchestration, multimodal, multi-turn dialogue) audited via traces and snapshots. Rubric scores with Pass^3 consistency. Claude Sonnet 4.6 81.4%
- DeepSearchQA In use Questions that need multi-step search and synthesis. Final answers graded against references. Kimi K3 95%