Small model. Fast function and tool selection.
Built for the decisions that turn AI into action.
Function calling is a core part of how AI agents get things done. To search for information, retrieve customer data, or take action in an application, an agent must choose the right function and provide the right arguments. MCP servers make tools available through a shared interface, but the agent still needs to decide what to call, when to call it, and when no call is needed.
Those decisions happen repeatedly. Using a large general-purpose model for every selection can add latency and cost before the actual work even begins.
That's why we built zlm-v1-tool-decisions.
ZeroGPU's 0.6B language model combines function selection, MCP tool routing, and argument generation in a single inference step. Built to run on the ZeroGPU edge network, it gives agents a compact, specialized model for choosing their next action.
Latency results below were measured on a cloud-hosted deployment, not edge devices.
Choose the function. Select the tool. Generate the arguments.
zlm-v1-tool-decisions reads the conversation and available function or tool descriptions, then generates the selection and inputs needed for the next step.
Function selection
Choose the appropriate function for the user's request.
MCP tool routing
Select tools exposed by MCP servers.
Argument generation
Produce search queries and structured inputs for the selected tools.
Context-aware decisions
Use conversation history and earlier tool results to determine what comes next.
No-call decisions
Return NO_TOOL when the request needs no external action or existing results already answer it.
The model makes the decision. Your agent executes the selected function or tool.
A specialized model for a focused job.
Evaluation 1, new questions: on 1,334 validation questions not present in training, zlm-v1-tool-decisions achieved 80.4% tool-choice accuracy. A correct result requires selecting the expected set of tools, including choosing no tool when appropriate.
Evaluation 2, blind reference: a separate blind-reference evaluation compared models on 106 clear routing examples.
| Model | Tool-choice accuracy |
|---|---|
| zlm-v1-tool-decisions | 83.0% |
| Qwen3-30B-A3B | 77.4% |
| GPT-6 Luna · High reasoning | 73.6% |
| GPT-6 Luna · Medium reasoning | 64.2% |
| GPT-6 Luna · No reasoning | 30.2% |
ZeroGPU recorded the highest observed accuracy in this sample, with fewer unnecessary calls than GPT-6 Luna at each tested reasoning level.
The blind reference was established by two AI judges. The sample is small, and the lead over Qwen3-30B and GPT-6 Luna high was not statistically significant. These scores measure tool selection, not end-to-end task completion.
Spend less time choosing the next action.
Function and tool selection sit directly in an agent's response path. Faster decisions help the agent reach the actual work sooner.
In the November 4 benchmark, both models were tested through the ZeroGPU API from the same client, with 238 sequential requests per model.
| Model | Median latency | p95 latency |
|---|---|---|
| zlm-v1-tool-decisions | 1,208 ms | 1,813 ms |
| GPT-6 Luna · High reasoning | 2,676 ms | 5,902 ms |
2.2× lower median latency. 3.3× lower p95 latency.
Additional API benchmark · October 27 · Virginia client
| Model | Median latency |
|---|---|
| zlm-v1-tool-decisions | 607 ms |
| GPT-6 Luna · No reasoning | 1,782 ms |
| GPT-6 Luna · Medium reasoning | 2,256 ms |
Results reflect steady-state cloud inference. Compare models within each test run; client location and network path affect latency.
Low token pricing for the calls agents make repeatedly.
At 1,000 input tokens and 20 output tokens per request, that is approximately $53 per million function-selection decisions.
Actual cost depends on conversation length, function descriptions, and generated arguments. Downstream tool execution is additional.
Give your agent a faster next step.
Function selection. MCP tool routing. Argument generation.
One compact model, purpose-built to turn intent into action.