HOST DIAGNOSTIC · Other
Kimi K2.6
Kimi K2.6· Other· OPEN WEIGHTS
ACTIVE
HOST IDENTIFICATION
Other
Kimi K2.6
ACTIVE
AVG SCORE
[10.7]
CAPABLE
STRONGEST
Planning & Execution
[13]
WEAKEST
Speed
[8]
MAXED (20/20)
0
of 20 attributes
20-DIMENSION EVALUATION
ATTRIBUTE SCORES · 20 DIMENSIONS
Cognitive
Bulk Apperception
11
Reasoning
11
Mathematical Precision
10
World Knowledge
10
Scientific Acumen
10
Technical
Code Generation
12
Tool Use
12
Multimodal Fluency
10
Speed
8
Cost Efficiency
13
Behavioral
Candor
9
Creativity
9
Tenacity
12
Self-Correction
11
Calibration
9
Operational
Instruction Following
11
Context Fidelity
11
Multi-turn Coherence
12
Planning and Execution
13
Safety Alignment
9
SCORING METHODOLOGY
CALIBRATED. Moonshot AI (China), released 2026-04-20. Open-weight 1T-param MoE / 32B active, 256K context, native multimodal (text+image+video), thinking + instant modes. CODING/AGENTIC SPECIALIST, not a universal frontier model. Independent sources: Artificial Analysis Intelligence Index v4.0 = 54 (#4 overall, #1 open-weight; the 57-tier is Anthropic/Google/OpenAI). GDPval-AA Elo 1520 (up from K2.5's 1309). tau2-Bench Telecom 96%. SWE-Bench Pro 58.6% (leads GPT-5.4 57.7%, Opus 4.6 53.4%, Gemini 3.1 Pro 54.2% under AA-matched max-reasoning conditions; lead is inside benchmark noise). SWE-Bench Verified 80.2% (Moonshot 10-run avg; trails Opus 4.7 87.6%). DIMENSION RATIONALE: Code Gen 17 + Planning/Exec 18 + Tool Use 17 + Tenacity 17 = agentic/long-horizon strength (300-sub-agent Agent Swarm, 13-hr autonomous runs, 1000+ tool calls, 4000+ lines modified). Speed 11 (DOWN from K2/K2.5's 14) = AA 47 tok/s, 'notably slow' 65/100, ~160M reasoning tokens for full Index, compute-heavy at full thinking. Cost Eff 18 = $0.60/$2.50 per M (~80% cheaper than closed frontier). Multimodal 14 = now native (K2.5 was 12). General-knowledge dims (Math/Sci/World 14) held near K2.5 — mid-tier, no standout gains. CAVEATS: Terminal-Bench 2.0 66.7% is Terminus-2-harness-specific — NOT a clean cross-model number. Independent benchmarking still thin (Epoch AI, METR, Stanford CRFM HELM have not published). Kilo FlowGraph practical test: K2.6 68/100 vs Opus 4.7 91/100 (multi-agent contention bugs). Anchored between Kimi K2 and Kimi K2.5 records. Sibling Kimi K2.7-Code (~2026-06-11) deferred — vendor-only proprietary benchmarks, fails traceable-source requirement.