switchboard-agent-router-0.5b-GGUF

The 994 MB F16 build of switchboardllm/switchboard-agent-router-0.5b, for Ollama and llama.cpp. It's a 0.5B router for agent loops. Before each step it answers which tier of model can make the agent's next tool call: {"tier":"small"}, "mid" or "frontier". No GPU, account or API key needed.

Run it

The easiest way is the switchboard-router package (0.7.0 or later). It pulls this model into Ollama the first time:

pip install switchboard-router

from switchboard_router import StepRouter

router = StepRouter(frontier_threshold=0.2)
router.pick_model(messages, tools)  # OpenAI-style chat messages and tool schemas

The model reads the compact text made by switchboard_router.render_step(messages, tools), not raw messages. To call it yourself, send that text as the single user message:

ollama pull hf.co/switchboardllm/switchboard-agent-router-0.5b-GGUF

import json, urllib.request
from switchboard_router import render_step

body = {"model": "hf.co/switchboardllm/switchboard-agent-router-0.5b-GGUF", "stream": False,
        "messages": [{"role": "user", "content": render_step(messages, tools)}]}
req = urllib.request.Request("http://127.0.0.1:11434/api/chat", json.dumps(body).encode(),
                             {"Content-Type": "application/json"})
print(json.load(urllib.request.urlopen(req))["message"]["content"])  # {"tier":"small"}

Or download the files and use the included Modelfile:

hf download switchboardllm/switchboard-agent-router-0.5b-GGUF --local-dir switchboard-agent
cd switchboard-agent
ollama create switchboard-agent -f Modelfile

With llama.cpp: llama-server -hf switchboardllm/switchboard-agent-router-0.5b-GGUF --temp 0 -n 24 -c 4096.

How close to the full model

On 300 held-out test steps, this build picked the same tier as the BF16 checkpoint on 288. The other 12 were near-ties. The median gap between their tier probabilities was 0.009.

Results

The full results are in the main model card. The test set is 2,805 held-out steps from tasks never seen in training:

• With the model's own pick, 13.5% were sent too low, saving 31.4% vs. always frontier.
• With frontier_threshold=0.2, 5.8% were sent too low, saving 19.2%.

If the output isn't valid JSON, use frontier. Don't use it as a security boundary: tool output and user text can try to talk it into a cheaper tier.

License

Apache-2.0, inherited from Qwen2.5-0.5B-Instruct. Training data includes nvidia/Nemotron-RL-Agentic-Function-Calling-Pivot-v1 by NVIDIA, used under CC-BY-4.0.