NextTokenAI/NextSearch-1-M

1---2license: apache-2.03base_model: thinkingmachines/Inkling-Small4pipeline_tag: text-generation5library_name: transformers6tags:7- agent8- web-research9- agentic-search10- tool-use11---12 13# NextSearch-1-M14 15NextSearch-1-M is the largest of the three **NextSearch-1** web research16agents: post-trained models that decompose a question, search and fetch from17the live web, reconcile conflicting evidence, and return a concise answer or18a structured research artifact. They are built to work as the research19component inside a larger system — called repeatedly by an orchestrator —20where per-call accuracy, tail latency, and cost compound.21 22| Model | Base | Params | |23|---|---|---|---|24| **NextSearch-1-M** (this repo) | Inkling-Small | 276B-A12B MoE | [weights](https://huggingface.co/NextTokenAI/NextSearch-1-M) |25| NextSearch-1-S | Qwen3.6-35B-A3B | 35B-A3B MoE | [weights](https://huggingface.co/NextTokenAI/NextSearch-1-S) |26| NextSearch-1-XS | Qwen3.5-9B | 9B dense | [weights](https://huggingface.co/NextTokenAI/NextSearch-1-XS) |27 28Technical report: **[nexttoken.co/research/nextsearch-1](https://nexttoken.co/research/nextsearch-1)**.29Harness, evaluation suite, and audited benchmark golds:30**[github.com/NextTokenAI/nextsearch](https://github.com/NextTokenAI/nextsearch)**.31 32## Results33 34Live-web evaluation (August 2026), 12B-active M against frontier API35anchors. Benchmarks: SEAL-0 (fresh/conflicting evidence, n=97), FRAMES36(multi-constraint retrieval, n=100), DeepSearchQA (comprehensive answer37sets, n=100), WideSearch-sub (structured table sub-tasks, n=49), and38WideSearch (full tasks under the orchestrated harness, n=20). Best per39column in **bold**.40 41| | SEAL-0 | FRAMES | DeepSearchQA | WideSearch-sub | WideSearch | mean $/ep | mean turns |42|---|---|---|---|---|---|---|---|43| **NextSearch-1-M** | **0.515** | 0.850 | **0.803** | 0.805 | 0.708 | $0.074 | 5.9 |44| glm-5.2 (355B-A32B) | 0.505 | **0.920** | 0.790 | 0.856 | **0.806** | $0.015 | 7.5 |45| gemini-3.6-flash | 0.495 | 0.880 | 0.773 | **0.885** | 0.712 | $0.127 | 8.8 |46| gpt-5.6-luna-med | 0.484 | 0.820 | 0.788 | 0.763 | 0.742 | $0.010 | 6.9 |47| deepseek-v4-flash | 0.474 | 0.820 | 0.781 | 0.682 | 0.758 | $0.029 | 10.1 |48| nemotron-3-ultra (550B-A55B) | 0.423 | 0.850 | 0.693 | 0.760 | — | $0.083 | 9.3 |49 50All rows run under **our harness** (same tools, prompts, turn budgets,51pinned task date) against **audited golds** with one shared judge —52consistent within this table, not comparable to other papers'53leaderboards. Protocol, costs, and reproduction:54[docs/evals.md](https://github.com/NextTokenAI/nextsearch/blob/main/docs/evals.md);55full analysis in the56[technical report](https://nexttoken.co/research/nextsearch-1).57 58## Quick start59 60The weights are ~530 GB bf16 — plan for a multi-GPU node (e.g. 8×H200).61See the [vLLM recipe for Inkling](https://recipes.vllm.ai/thinkingmachines/Inkling)62for current serving flags; our serving notes (sampling, context caps,63tool-call parsing pitfalls) are in64[docs/serving.md](https://github.com/NextTokenAI/nextsearch/blob/main/docs/serving.md).65 66```bash67vllm serve NextTokenAI/NextSearch-1-M --tensor-parallel-size 8 \68  --enable-auto-tool-choice --max-model-len 6553669```70 71Recommended sampling: temperature 0.7, max 16k tokens per turn, reasoning72effort 0.7. The model expects a task date in its system prompt and two73tools (`search`, `fetch`); the exact prompts and tool schemas it was74tuned for ship in the [harness](https://github.com/NextTokenAI/nextsearch):75 76```bash77pip install nextsearch && nextsearch-eval run --benches seal0 --models nextsearch-1-m --n 1078```79 80## License81 82Released under the **Apache License 2.0**, as is the base model83[`thinkingmachines/Inkling-Small`](https://huggingface.co/thinkingmachines/Inkling-Small).84 85## Citation86 87```bibtex88@techreport{nextsearch1,89  title       = {NextSearch-1: Open models for wide and deep web research},90  author      = {Nitish Kulkarni and Alankar Jain},91  institution = {NextToken},92  year        = {2026},93  url         = {https://nexttoken.co/research/nextsearch-1}94}95```96