Moonshot AI's October 23 release. BenchCAD 92.7, financial-research hallucination cut to 8.3%. Vulnerability-research data stripped before open-weight release.
Parameters
TBA
Context Window
TBA
License
Open-weight (non-commercial)
Release Date
2026-10-23
API Pricing
API pricing for this model is not yet available
Strengths
- •Financial-research hallucination 8.3% (was 23.4% on K3)
- •Inspected the simulator's own errors
- •Stripped vulnerability data before open-weight release
Weaknesses
- •BenchCAD is a Moonshot-shaped eval
- •81% exploit-solve is not small
- •No context window or price published
Use Cases
- •Finance, law and engineering draft review
- •Professional work where hallucination costs real money
- •Open-weight deployment
Deep Analysis
BenchCAD
92.7
OfficeQA Pro v2
64.4
Hallucination (finance)
8.3% (was 23.4% on K3)
Open-weight
Yes, vuln-research data stripped
Strengths
- ・Real professional-delivery signal: financial-research hallucination dropped from 23.4% on K3 to 8.3%, the number that decides whether a report is safe to ship.
- ・It inspected the simulator itself — found 36 errors in shoulder-muscle math — instead of grinding inside broken rules. That is the behavior I actually want from a frontier model.
- ・Moonshot stripped vulnerability-research data before open-weight release (exploit-solve 90.5% to 81%), choosing not to push offensive frontier into the open.
Weaknesses
- ・BenchCAD and OfficeQA are Moonshot-shaped evals; we need independent runs before trusting the 92.7.
- ・Open-weight with stripped safety data still ships real capability; the 81% exploit-solve is not small.
- ・No context window, price, or standard leaderboard numbers published yet.
Competitor Comparison
| Model | GPQA | Price |
|---|---|---|
| DeepSeek V4.1-Flash | TBA | open |
| Qwen3.5-397B-A17B | TBA | open |
| GPT-5.6 Sol | 94.6% | $5/$30 |
Kimi K3.1 Pro is Moonshot AI's October 23, 2026 release. In internal robotics-sim tests the model stood 11.2 seconds (up from falling over in seconds) and flagged 36 errors in the simulator's own shoulder-muscle math. On BenchCAD it scored 92.7 and cut financial-research hallucination to 8.3% from K3's 23.4%. Moonshot removed vulnerability-research data before open-weight release.
Analysis generated: 2026-10-27