NextSearch-1-XS
NextSearch-1-XS is the smallest NextSearch-1 web research agent: a post-trained model that decomposes a question, searches and fetches from the live web, reconciles conflicting evidence, and returns a concise answer or a structured research artifact. The family is built to work as the research component inside a larger system — called repeatedly by an orchestrator — where per-call accuracy, tail latency, and cost compound. XS is the price/latency point: a 9B dense model you can serve on one GPU.
| Model | Base | Params | |
|---|---|---|---|
| NextSearch-1-M | Inkling-Small | 276B-A12B MoE | weights |
| NextSearch-1-S | Qwen3.6-35B-A3B | 35B-A3B MoE | weights |
| NextSearch-1-XS (this repo) | Qwen3.5-9B | 9B dense | weights |
Technical report: nexttoken.co/research/nextsearch-1. Harness, evaluation suite, and audited benchmark golds: github.com/NextTokenAI/nextsearch.
Results
Live-web evaluation (August 2026), against small and cheap models. Benchmarks: SEAL-0 (fresh/conflicting evidence, n=97), FRAMES (multi-constraint retrieval, n=100), DeepSearchQA (comprehensive answer sets, n=100), WideSearch-sub (structured table sub-tasks, n=49). Best per column in bold.
| SEAL-0 | FRAMES | DeepSearchQA | WideSearch-sub | mean $/ep | mean turns | |
|---|---|---|---|---|---|---|
| NextSearch-1-XS | 0.289 | 0.790 | 0.588 | 0.703 | $0.093 | 7.0 |
| claude-haiku-4.5 | 0.258 | 0.720 | 0.670 | 0.864 | $0.154 | 8.7 |
| qwen3.5-9b (base) | 0.206 | 0.700 | 0.599 | 0.610 | $0.030 | 9.7 |
| gemma-4-31b | 0.155 | 0.670 | 0.565 | 0.667 | $0.013 | 5.0 |
| gemini-3.5-flash-lite | 0.216 | 0.570 | 0.535 | 0.640 | $0.020 | 7.3 |
| gpt-oss-20b | 0.206 | 0.710 | 0.406 | 0.470 | $0.012 | 10.5 |
The table above is the conservative arm (parallel search backend). With the recommended exa-auto backend XS's four-bench mean rises from 0.592 to 0.693 (0.412 / 0.860 / 0.771 / 0.728) — above every genuinely small model on every bench except haiku's WideSearch-sub, and edging the 550B nemotron-3-ultra anchor (0.682) at a fraction of the size.
All rows run under our harness (same tools, prompts, turn budgets, pinned task date) against audited golds with one shared judge — consistent within this table, not comparable to other papers' leaderboards. Protocol and reproduction: docs/evals.md; full analysis in the technical report.
Quick start
vllm serve NextTokenAI/NextSearch-1-XS \
--enable-auto-tool-choice --tool-call-parser hermes --max-model-len 65536
Recommended sampling: temperature 0.7, max 16k tokens per turn, thinking
on. The model expects a task date in its system prompt and two tools
(search, fetch); the exact prompts and tool schemas it was tuned for
ship in the harness:
pip install nextsearch && nextsearch-eval run --benches seal0 --models nextsearch-1-xs --n 10
Or plain transformers:
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"NextTokenAI/NextSearch-1-XS", torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("NextTokenAI/NextSearch-1-XS")
Serving pitfalls that fail silently (tool-call parsing, context caps, thinking retention): docs/serving.md.
License
Released under the Apache License 2.0, as is the base model
Qwen/Qwen3.5-9B.
Citation
@techreport{nextsearch1,
title = {NextSearch-1: Open models for wide and deep web research},
author = {Nitish Kulkarni and Alankar Jain},
institution = {NextToken},
year = {2026},
url = {https://nexttoken.co/research/nextsearch-1}
}