Answers from the documentation; checking which model is on duty
Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
SID-2 Agentic Search Model Launches with Swarm Mode and 0.720 NDCG - OpenSmartRoute
Introduction - SID-2 release and its significance in search models
SID AI has launched SID-2, its second-generation agentic search model. This new model introduces swarm mode, which significantly improves retrieval quality. It marks a step forward from previous search models by using multiple agents to explore data more effectively. The release highlights ongoing progress in making search models faster and more accurate.
This launch is important because it replaces older fixed pipelines with a more dynamic approach. Traditional search systems rely on static processes that do not adapt during a search. SID-2 changes this by allowing the model to read, decide what to search next, and adapt its strategy. This makes search more flexible and capable of handling complex data sets.
Model overview - architecture, size, and deployment options
SID-2 is built with a large context window, capable of processing up to 262,000 tokens. This allows it to handle extensive data and complex queries without losing context. The model is designed as a small search specialist, not a general-purpose chat model. It focuses solely on searching, reading results, and returning ranked document IDs.
Developers can deploy SID-2 through various options. It is available via SID’s hosted API at platform.sid.ai. It can also be integrated with Baseten and SageMaker, which are platforms for deploying machine learning models. The model uses open-weight models that can be served locally or through cloud providers. Its deployment options give teams flexibility depending on their infrastructure needs.
Performance metrics - benchmark results and comparisons
SID-2 has been tested on a benchmark with 506 questions across 19 different data sources. These sources include legal documents, emails, scientific articles, and more. The results show that SID-2 performs well compared to other models.
A single SID-2 agent with medium reasoning effort scores 0.585 NDCG, a common measure of retrieval quality. This is comparable to Claude Opus 5.5, which scores 0.577. When using swarm mode with six agents, SID-2 reaches 0.701 NDCG in less time. With twelve agents, it scores 0.720 NDCG, the highest in the test. These results show that more agents improve retrieval quality significantly.
Claude Fable 5.2 creates images by simulating brushstrokes through Python code, mimicking physical painting. It can produce complex, photorealistic art and plan physical tasks.
SID-2’s best performance surpasses models like GPT-6 Astra, which scores 0.704 NDCG. It also beats traditional search methods like BM25, which scores 0.260. The results demonstrate that SID-2 can deliver high-quality search results faster and at lower cost than some competitors.
Swarm mode - how multiple agents explore data slices
Swarm mode involves multiple SID-2 agents working together. Each agent explores a different part of the data corpus. They communicate and share information during the search process. This approach allows the system to read more documents and gather more relevant information.
The idea behind swarm mode is that reading more documents improves search quality more than reasoning harder about each document. Each agent forks off into a new thread, exploring a separate slice of data. This parallel exploration speeds up the process and increases the chances of finding the best results.
SID argues that swarm mode is more efficient than a single agent trying to reason about all data at once. It allows the system to cover more ground in less time. The number of agents can be adjusted, with more agents generally leading to better results, but also higher costs. The 12-agent swarm achieves the highest retrieval score, showing the potential of this approach.
Training process - reinforcement learning stages and optimization
SID-2 was trained in two main stages using reinforcement learning. The first stage focused on optimizing retrieval quality. It aimed to improve how well the model finds relevant documents. The second stage added a latency penalty, which discourages slow responses.
During training, SID used a self-play loop. It generated hard queries against a frozen version of SID-1, an earlier model. This process created difficult questions that pushed SID-2 to improve. The training setup achieved a sixfold increase in throughput compared to previous methods.
The training process also involved merging different levels of reasoning effort. SID-2 was trained with low, medium, and high reasoning effort as separate experts. These were combined through multi-teacher on-policy distillation, which helps the model learn from multiple strategies. This approach resulted in a versatile model that can adapt its reasoning effort based on user needs.
Benchmark details - datasets, scores, and evaluation results
The benchmark used for SID-2 includes a diverse set of datasets. These cover areas like question answering, emails, legal filings, scientific literature, and more. The goal was to test retrieval quality across different types of data.
SID-2’s medium effort version scored 0.585 NDCG in 3.11 seconds per query. This is comparable to Claude Opus 5.5, which scored 0.577 in 25.1 seconds. The swarm mode with six agents scored 0.701 NDCG in 17.7 seconds, while twelve agents reached 0.720 NDCG. These scores show that SID-2 can deliver high-quality results quickly.
The results also include traditional search baselines. BM25, a keyword-based method, scored 0.260, and a vector search with reranking scored 0.287. SID-2’s scores are significantly higher, especially in swarm mode, demonstrating its advanced retrieval capabilities.
Pricing and cost considerations - API costs and efficiency
SID-2 is priced at $0.25 per million input tokens and $1.00 per million output tokens. Swarm mode, which uses multiple agents, costs more—$0.50 per million input tokens and $2.00 per million output tokens. Cached input, which avoids reprocessing, costs only 10% of the input price.
The cost efficiency of SID-2 depends on the number of agents and the reasoning effort level. A single high-effort agent can match or beat the performance of multiple SID-1 runs at a fraction of the cost. Swarm mode offers the best retrieval quality but at higher expense. Teams can choose the level of effort and number of agents based on their budget and accuracy needs.
Because SID-2 does not require reindexing or data movement, it can be integrated into existing workflows without additional data management costs. This makes it suitable for teams seeking high-quality search without significant infrastructure changes.
Background - what agentic search models replace and how they work
Traditional search systems often use fixed pipelines. These pipelines typically involve rewriting queries, embedding them into vectors, searching, and reranking results. These processes are static and do not adapt during a search.
Agentic search models like SID-2 replace these fixed pipelines with models that make decisions during the search process. Instead of following a set order, SID-2 searches, reads, and adapts based on what it finds. It can decide what to search next, improving relevance and speed.
This approach allows for more flexible and intelligent search. It can handle complex queries and large data sets more effectively. It also reduces the need for reindexing or moving data, as the model only sees what tools return during each episode.
Why it matters - implications for search quality and speed
The launch of SID-2 and its swarm mode has several implications. First, it improves retrieval quality, as shown by the high NDCG scores. Better retrieval means users get more relevant results faster. This can enhance decision-making and knowledge discovery.
Second, the model’s ability to explore data slices in parallel speeds up the search process. Swarm mode reduces response times significantly compared to traditional methods. Faster search can improve user experience and support real-time applications.
Third, the flexible training and deployment options make SID-2 adaptable to different needs. Teams can tune the reasoning effort and cost to match their priorities. This flexibility allows for more efficient resource use and better results.
Overall, SID-2 pushes the boundaries of what search models can do, making them more accurate and faster. It offers a new way to handle large, complex data sets with fewer resources.
How OpenSmartRoute helps
Teams routing requests through OpenSmartRoute can benefit from SID-2’s advancements in search quality and efficiency. OpenSmartRoute’s platform scores each candidate model or tool on quality, cost, speed, and safety. It allows teams to set weights based on their priorities.
Using SID-2 with OpenSmartRoute, teams can route queries to the most suitable model or swarm configuration. They can also enforce rules, such as keeping personal data on on-premises models or staying within region limits. The platform learns from outcomes, directing more traffic to models that perform well.
OpenSmartRoute’s catalog of models and tools makes it easy to integrate SID-2 into existing workflows. Its savings ledger shows the cost of each routed request compared to the most expensive option. This setup helps teams optimize search quality and cost over time.
How it compares - what existed before, what this changes and what stays the same
Before SID-2, most search systems used fixed pipelines. These pipelines followed a set order of steps, such as rewriting, embedding, searching, and reranking. Each step was separate, and the process was static. Developers had to reindex data whenever updates occurred.
Traditional retrieval models relied on keyword matching or vector search. These methods were simple and fast but often less accurate. They did not adapt during search and could not decide what to search next based on results.
SID-2 changes this by replacing fixed pipelines with an agentic approach. Instead of following a set order, SID-2 searches, reads, and adapts during each episode. It decides what to search next based on previous results. This makes the process more flexible and intelligent.
SID-2 also introduces swarm mode, where multiple agents explore different parts of the data simultaneously. This parallel exploration improves retrieval quality and reduces response times. Swarm mode can outperform traditional models on many datasets.
Despite these changes, some aspects stay the same. The core goal remains to find relevant documents quickly. The tools used for keyword and vector search are still supported. Developers can integrate SID-2 with their existing backends without reindexing data.
In summary, SID-2 advances search by making it more dynamic and scalable. It replaces static pipelines with adaptive agents and parallel exploration, leading to better results and faster responses.
Questions this leaves open - what the source does not say and how a reader can check it
The source does not specify the exact size or architecture of SID-2. It mentions the context window size but does not disclose the model’s total number of parameters or layers. Without these details, it is hard to compare SID-2’s capacity to other models.
It also does not provide independent evaluation results. All numbers are self-reported from SID’s benchmark. External tests could reveal how SID-2 performs on unseen data or in real-world scenarios. Readers can check for independent benchmarks or run their own tests if possible.
The source does not clarify how SID-2 handles data privacy or security. Since it only sees what tools return, it may be easier to control sensitive information. However, details about data handling and compliance are missing. Users should review SID-2’s deployment options and security features.
Pricing details are limited. The report states costs per million tokens but does not explain how these compare to other search solutions or large language models. Teams can estimate costs based on their expected query volume and compare with existing tools.
The report mentions that SID-2 is available via API and SDKs, but it does not specify the level of customization or fine-tuning possible. Users interested in domain-specific tuning should inquire about options and limitations.
Finally, the report does not describe how SID-2’s reasoning effort levels are chosen or adjusted in practice. It mentions training separate experts for each level, but not how users select or switch between them during deployment. Teams can experiment with different settings to find the best balance of quality and speed.
In conclusion, while SID-2 offers promising improvements, some details remain unclear. Developers and managers should seek additional information through technical documentation, independent evaluations, or direct testing. This will help them understand the model’s capabilities and limitations better.
Google introduced Gemini 4 Hydrogen, a faster model for robotics that combines strong spatial reasoning with efficiency. It supports real-world robot control and simulation.