Ultra mode (4 parallel agents) sets BrowseComp SOTA at 92.2%.

Context compaction at 300K tokens. 90.4 with no context mgmt.

Beats Mythos 5 (88.0%) and Fable 5 (84.3%); Ultra mode hits 92.2%.

First open 3T-class model (2.8T params). Score achieved with 1M-token context window and no context management (vs Claude-style compaction).

New BrowseComp SOTA. Nearly 5pt over Gemini 3.1 Pro.

Highest BrowseComp score reported. Pro variant with maximum reasoning.

Above Mythos Preview (87.9) and Fable 5 (84.3); below GPT-5.6 Sol (90.4) and Sol Ultra (92.2).

Score as cited in Kimi K3 blog (source: Anthropic official). Fable 5 hit fallbacks on 35% of tasks.

Terra surpasses GPT-5.5 (84.4%) on BrowseComp at lower cost.

New SOTA. Uses 4.9x fewer tokens than Opus 4.6 while scoring higher.

Agent Swarm mode with parallel sub-agents. 2nd highest BrowseComp score after GPT-5.4 Pro.

Up from 59.2% on Gemini 3 Pro. Strong autonomous web research capability.

New BrowseComp SOTA for autonomous research. Uses Gemini 3.1 Pro with extended search and reasoning.

Beats GPT-5.4 (82.7) but trails Gemini 3.1 Pro (85.9) and GPT-5.5 Pro (90.1).

Score as cited in Kimi K3 blog (source: Anthropic official).

Multi-agent web search coordination across hours-long sessions. 86.8% with harness.

Max reasoning. Competitive with GPT-5.4 (82.7), below GPT-5.5 Pro (90.1).

Luna nearly matches GPT-5.5 peak (84.4%) at less than half the cost.

Open-weight 1T/32B MoE. Strong web research score, beating GPT-5.4 at 82.7%.

Up from 65.8% on GPT-5.2. Native computer-use model with 1M context.

With context management. No-tools version: 68.0%. Strong browsing capabilities.

Notable regression from Opus 4.6 (83.7%). First major BrowseComp decline in a frontier model update.

With aggressive context-folding strategy. Outperforms every US frontier model on web browsing.

With Agent Swarm. Base score 60.6, with context management 74.9. Competitive with GPT-5.4 (82.7).

Multi-agent swarm configuration. 72.7% on WideSearch.

Strong browsing for open-weights; below GPT-5.6 Sol Ultra (92.2) but above Kimi K2.5 (74.9).

Strong browsing agent performance. Competitive with Claude Opus 4.6 (84.0%).

With context management strategy. Baseline without CM: 62.0%. Highest open-model BrowseComp score reported.

BrowseComp. Approaches Claude Opus 4.7 (79.3) at Flash scale.

Single-agent config. Corrected after cheating detection pipeline update (was 74.72%, adjusted to 74.01%). Multi-agent variant scores 82.07%.

Max reasoning, 13B activated. Respectable for small model.

Score with Context Manager agent framework. Base score 51.6. Strong browsing for an open model.

Top open-model score on BrowseComp. 79.3 with context management variant.

Small model (27B params) with competitive browsing ability. Below frontier but strong for open-weight.

Single-agent mode. Agent Swarm achieves 78.4%.

First fully open-source search agent achieving frontier-level performance. OpenSeeker-30B trained on only 11.7k samples achieves 29.5% on BrowseComp and 48.4% on BrowseComp-ZH, surpassing industrial c

India's first domestically-trained 105B model. MoE with 10.3B active params. Apache 2.0.

First fully open-source search agent achieving frontier-level performance. OpenSeeker-30B trained on only 11.7k samples achieves 29.5% on BrowseComp and 48.4% on BrowseComp-ZH, surpassing industrial c

BrowseComp baseline (no context management).

Lightweight flash variant browser capability score.

First fully open-source search agent achieving frontier-level performance. OpenSeeker-30B trained on only 11.7k samples achieves 29.5% on BrowseComp and 48.4% on BrowseComp-ZH, surpassing industrial c

First fully open-source search agent achieving frontier-level performance. OpenSeeker-30B trained on only 11.7k samples achieves 29.5% on BrowseComp and 48.4% on BrowseComp-ZH, surpassing industrial c

8B deep research agent with verification-centric design surpasses 30B-scale agents on BrowseComp. RL stage adds +6.7 on xBench-DeepSearch.

8B deep research agent with verification-centric design surpasses 30B-scale agents on BrowseComp. RL stage adds +6.7 on xBench-DeepSearch.

First fully open-source search agent achieving frontier-level performance. OpenSeeker-30B trained on only 11.7k samples achieves 29.5% on BrowseComp and 48.4% on BrowseComp-ZH, surpassing industrial c