Source retrieval
Connect to a controlled search or corpus index and record the precise documents returned for each query.
AGENT USE CASE · SOURCE-FIRST
Research agents.
Research-agent quality depends on retrieval, source selection, faithful citations and multi-step execution, not just the language model. Model-only reasoning benchmarks cannot substitute for end-to-end research evaluations.
What must be evaluated?
Citation grounding
Require verifiable passages and distinguish primary evidence from model-generated explanation.
Planning and stopping
Test follow-up searches, contradiction handling, context limits and explicit abstention when evidence is absent.
System evaluation
Report factual support, coverage and time/cost on held-out research tasks. Do not transfer benchmark results between agents.
21 catalog candidates and components
These model records are included by catalog modality or category, not by a verified task-success benchmark. Follow their primary sources before testing.
MiniMax-M3Core candidate · MiniMax Community LicenseInspect model →K2-Horizon-375B-A23BCore candidate · apache-2.0Inspect model →MiMo-V2.6-Pro-RLCore candidate · mitInspect model →GLM-5.3Core candidate · otherInspect model →Kimi-K3Core candidate · otherInspect model →GLM-5.3-FlashCore candidate · mitInspect model →DeepSeek-V4.1-FlashCore candidate · mitInspect model →Qwen3.8-27BCore candidate · apache-2.0Inspect model →InklingCore candidate · apache-2.0Inspect model →NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16Core candidate · otherInspect model →Muse-Glimmer-30BCore candidate · apache-2.0Inspect model →gpt-oss-120bCore candidate · apache-2.0Inspect model →gemma-4-26B-A4B-itCore candidate · apache-2.0Inspect model →Mistral Small 4 119B A6BCore candidate · apache-2.0Inspect model →Olmo-3-7B-ThinkCore candidate · apache-2.0Inspect model →K2-Horizon-MoVA-36B-A4BCore candidate · apache-2.0Inspect model →clef-omniSupporting component · apache-2.0Inspect model →pplx-embed-v2-late-0.6bSupporting component · mitInspect model →pplx-embed-v2-late-9bSupporting component · mitInspect model →Darwin-27B-ZTCSupporting component · apache-2.0Inspect model →Granite 4.2 30BCore candidate · apache-2.0Inspect model →
COMMON QUESTIONS
Frequently asked questions
Does a long context window ensure good research?
No. Configured context length is not a measure of factuality, retrieval quality or citation precision.
Can research run privately?
Yes in principle when the model, search corpus and all connected services stay within a controlled boundary; every connector must be checked.
READ THE ORIGINALS
Documentation and primary sources
REAL SOURCE EVIDENCE
Publisher-reported task results
Different benchmark versions are not comparable; no numbers below were independently reproduced by Open Model Weights.
Mellum2.1-12B-A2.5B-Thinking · BFCL v462.3 percent · publisher-reportedReview protocol and source →K2-Horizon-375B-A23B · tau3-Banking34 reported score · publisher-reportedReview protocol and source →K2-Horizon-375B-A23B · Toolathlon Verified65.3 reported score · publisher-reportedReview protocol and source →K2-Horizon-375B-A23B · MCPMark67.7 reported score · publisher-reportedReview protocol and source →
Evaluate before deployment.
Model metadata is source-linked; agent task success, tool-call reliability and production readiness require separate tests.
Review the methodology → Return to the explorer →