Skip to content
Other

OpenToolSearch Reranker: A compact reranker for selecting AI agent tools

3 min readupdated

This model from AI3 Inc. helps AI agents efficiently handle massive libraries of functions and tools (such as function declarations or MCP tools) without overloading their system prompts. OpenToolSearch Reranker works as a second-stage search: a 0.6B-parameter cross-encoder (a modified causal LLM with a yes/no output) reads the query alongside the tool description and scores its relevance. Built on Qwen/Qwen3-Reranker-0.6B, this solution paired with a first-stage embedder outperforms many heavier alternatives while saving tokens and response time.

Key features

  • Precise selection: A 0.6B cross-encoder evaluates the match between the query and tools on a binary scale (yes/no).
  • Balanced architecture: Weights were averaged 50/50 with the base model, and JSON and YAML template augmentation was applied during training.
  • Resource efficiency: Works alongside a fast first-stage embedder (such as nvidia/Nemotron-3-Embed-1B-BF16) to filter out excess tools before calling the main LLM.
  • English optimization: Trained on cleaned ToolRet, Toucan-1.5M, and Glaive function-calling v2 datasets.

What you can do

  • Organize tool search: Create a local tool index for an AI agent to pass only the necessary functions during a conversation.
  • Accelerate agent pipeline: Use a two-stage search—first selecting top-50 candidates via an embedder, then reranking them with this model.
  • Integrate into custom frameworks: Use the model with the transformers library to build custom API search engines.

Example

Running and evaluating tool relevance using the transformers library:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

name = "ai3inc/opentoolsearch-rerank-0.6b"
tok = AutoTokenizer.from_pretrained(name, padding_side="left")
model = AutoModelForCausalLM.from_pretrained(name, torch_dtype=torch.bfloat16).eval()
yes, no = tok.convert_tokens_to_ids("yes"), tok.convert_tokens_to_ids(
    "no"
)

PRE = (
    "<|im_start|>system\nJudge whether the Document meets the requirements"
    " based on the Query and the Instruct provided. Note that the answer can"
    ' only be "yes" or "no".<|im_end|>\n<|im_start|>user\n'
)
SUF = "<|im_end|>\n<|im_start|>assistant\n\n\n\n\n"
DEFAULT = "Given a user request, retrieve the tools that are needed to fulfil it"

def score(request, tool_json, instruction=None):
  text = f"{PRE}: {instruction or DEFAULT}\n: {request}\n: {tool_json}{SUF}"
  ids = tok(text, return_tensors="pt", add_special_tokens=False).to(
      model.device
  )
  with torch.no_grad():
    logits = model(**ids).logits[0, -1]
  return (logits[yes] - logits[no]).item()

tool = '{"name": "get_weather_forecast", "description": "Get the weather forecast for a city on a given date."}'
print(score("What's the weather in Seoul tomorrow?", tool))
Read moreDetails
Benchmarks: On the ToolRet benchmark (ACL 2025), pairing with Nemotron-3-Embed-1B achieves 58.37 nDCG@10 (top-100 reranking), trailing larger alternatives by only about 1.5 points while being significantly smaller.
Hardware and VRAM: 0.6B parameters in bfloat16 format take up about 1.2 GB of VRAM.
Training: 300 steps of fine-tuning on a single NVIDIA B300 GPU (learning rate 6e-6), using hard negatives from the first-stage embedder.
Limitations: Only English is supported; requires 50–100 forward passes per search query (depending on the number of candidates).
Discuss Telegram 1 62

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 1.1 GB · ready files1.7 GB
  • FP8 / INT8recommendedweights 0.6 GB1.1 GB
  • 4-bitweights 0.3 GB0.8 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # ai3inc/opentoolsearch-rerank-0.6b
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download ai3inc/opentoolsearch-rerank-0.6b --local-dir ~/models/opentoolsearch-rerank-0.6b
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

  1. Qwen3-0.6B-Basefamily root
  2. Qwen3-Reranker-0.6B
  3. opentoolsearch-rerank-0.6b

New posts about versions of Qwen3-0.6B-Base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 3All on Hugging Face

About the model

ai3inc/opentoolsearch-rerank-0.6b
Author
ai3inc
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

621

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments