TL;DR
Moonshot’s Kimi K3 debuted at No. 3 on VigilSAR’s public defense-ISR LLM leaderboard, scoring 64.65 in Band B. The private-task benchmark placed Kimi K3 ahead of every listed GPT and Gemini model, although VigilSAR cautions readers to compare score bands rather than rank numbers.
Moonshot’s Kimi K3 debuted at No. 3 on VigilSAR’s public language-model leaderboard, scoring 64.65 in Band B on a private set of defense intelligence, surveillance and reconnaissance tasks. The result placed Kimi K3 ahead of every GPT and Gemini entry shown on the board as of July 17, 2026.
VigilSAR evaluated 14 language models across 300 tasks designed to test reasoning, reporting and restraint in defense-ISR work. Aggregate scores are public, but the tasks remain private to reduce the risk that model developers train directly on the evaluation material.
The leaderboard lists claude-fable-5 as the leader and pinned reference model, with a score of 67.77 in Band A. Kimi K3 follows in Band B at 64.65 and appears third in the displayed standings. The GPT-5.x family occupies Bands C and D, while the listed Gemini models sit in Bands E and F.
VigilSAR also uses a separate held-out task set and publishes the gap between each model’s public and held-out results as a possible warning of memorization. The board reports confidence intervals and cost per correct answer, linking capability scores with the cost of using each model.
Kimi Challenges Larger Model Families
Kimi K3’s placement gives buyers and developers another data point when comparing models for specialized analytical work. On this benchmark, it outscored all listed models from the GPT and Gemini families, suggesting that performance on broad consumer benchmarks may not predict results on constrained defense-ISR tasks.
The benchmark also treats deployment conditions as part of model evaluation. One locally runnable open model carries a “sovereign-deployable” label, reflecting the role of local control, security requirements and operating cost when models are considered for sensitive environments. The results are historical benchmark measurements, not guarantees of future performance or suitability for any deployment.
defense ISR language model
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Private Tasks Test Operational Restraint
VigilSAR is a defense-ISR software product that created the benchmark to compare models it may use in its own systems. Unlike general-knowledge tests, the evaluation focuses on the reasoning, reporting discipline and restraint expected from an analyst handling intelligence-related tasks.
The operators say they receive no payment from model vendors. Their published approach favors score bands over exact rank comparisons because confidence intervals for models within the same band overlap. The pinned reference row, held-out gaps and cost figures are intended to make changes across leaderboard updates easier to interpret.
“Vendor claims are not evidence.”
— VigilSAR benchmark page
DULIWO Scribing Tool Kit for Gunpla – 7-Blade Model Scriber Chisel Set (0.1–2.0mm), Pin Vise Hand Drill with 10 Bits, Tweezers & Brush Included for Gundam HG/RG/MG, Resin Kits, Panel Line Engraving
Model Kit Tools: Includes metal scribe tool (1 handle + 7 blades), small hand drill set (10 bits…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Rank Precision Remains Limited
Kimi K3’s No. 3 position is confirmed on the displayed leaderboard, but VigilSAR warns that readers should compare bands rather than rank numbers. Overlapping confidence intervals mean a numbered position may imply more separation than the data supports.
The private task design limits contamination, but it also means outsiders cannot inspect the individual prompts or expected answers. The available information does not establish how Kimi K3 performed in each task category, whether the benchmark has received an independent audit, or how stable its score would be across repeated evaluations.
private task AI models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Future Runs Will Test Durability
Later leaderboard updates will show whether Kimi K3 retains Band B as VigilSAR adds models or repeats testing. Changes in its held-out gap, confidence interval and cost per correct answer will offer a fuller view of whether the debut result holds under further evaluation.
Source: Thorsten Meyer AI
local deployment AI models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What score did Kimi K3 receive?
Kimi K3 scored 64.65, placing it in Band B and third in the displayed standings on July 17, 2026.
Did Kimi K3 beat GPT and Gemini models?
On this specific VigilSAR evaluation, Kimi K3 ranked ahead of every listed GPT and Gemini entry. That result applies to the benchmark’s 300 defense-ISR tasks and does not establish superiority across all uses.
Which model led the leaderboard?
claude-fable-5 led the published standings with 67.77 in Band A. VigilSAR also uses it as the pinned reference row.
Why are the benchmark tasks private?
VigilSAR keeps the task set private to reduce training contamination and direct memorization. It also uses a separate held-out set to compare model performance on material not represented in the public aggregate result.
Does third place mean Kimi K3 is definitively the third-best model?
No. It is the third-listed model in the current results, but VigilSAR says score bands and confidence intervals provide a more reliable comparison than exact rank numbers.
Source: Thorsten Meyer AI
