Is SID-2 Swarm state of the art?Pending benchmarks.Announced Nov 7, 2026; no independent results yet.
SID AI's multi-agent search model, kept separate from SID-2. Official docs: a root agent forks into a team that searches in parallel, messages other agents, and reports ranked document IDs, on the first-party SID API over a WebSocket (model SID-2-Swarm). Official models page: reasoning is fixed at high, context is described as effectively infinite, and list pricing is $0.50 input, $0.05 cached input, and $2.00 output per 1M tokens. Official technical report dated November 7, 2026, describes SID-2 Swarm as trained as a team of agents. Not an LLM for general chat.
The details so far
- Lab
- SID AI
- Announced
- Nov 7, 2026
- API
- Live · SID-2-Swarm
- Open weights
- Closed · Not released
- Task
- Multi-agent agentic search: a team of agents returns document IDs in relevance order that the caller can fuse, for example with RRF (official Swarm docs)
- Context
- Official models page: effectively infinite. Official Swarm docs: each agent can have its own context after a fork
- Reasoning
- Fixed at high; official request docs say any other effort value is refused
- Agents
- Official request docs: multi_agent.max_agents from 1 to 24, default 6, and not more than the account's agent quota (12 by default)
- Availability
- Public first-party SID API at wss://api.sid-1.com/v1/responses, model SID-2-Swarm, with the same self-serve SID API key as SID-2 (official Swarm docs). Official overview says it is currently only available via SID's first-party API
- Price
- $0.50 input, $0.05 cached input, and $2.00 output per 1M tokens (official SID models page)
- Related
- A separately trained multi-agent model related to SID-2, not only a SID-2 mode (official technical report and models page)
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: on SID's self-reported 506-question benchmark across 19 corpora, SID-2 Swarm with 12 agents scored 0.720 NDCG, and with 6 agents scored 0.701 NDCG with an estimated 17.7 seconds of model inference per query (official technical report, Tables 2 and 4). The same official tables list GPT-6 Astra at maximum effort at 0.704 NDCG and 135 seconds.
Sources
LABSID-2 Technical Report: Scaling Search Swarmssid.ai/research/sid-2-technical-report · Nov 7, 2026LABSID models: SID-2, SID-2 Swarm and SID-1sid.ai/modelsLABSID-2 Swarm overview (official docs)platform.sid.ai/docs/sid-2-swarmLABSID-2 Swarm making a request (official docs)platform.sid.ai/docs/sid-2-swarm/making-a-requestTracking continues until an independent board lists it.