Moonshot AI’s Kim K3.1 Pro is challenging Claude Opus 5.5 and GPT-6 Astra in financial analysis, while trailing both on a 3D modeling benchmark. Launched on October 23, 2026, the upgraded large language model has 3 trillion parameters and a one million token context window, with model weights planned for public release by November 4.
Key takeaways
- Moonshot reports a financial-analysis benchmark win over both rivals.
- Terminal-Bench 4.0 puts the model between Astra and Opus 5.5.
- Paid cloud access is live ahead of the planned weights release.
Moonshot AI’s Kim K3.1 Pro builds on selective activation
Kim K3.1 Pro upgrades Kimi K3, the model Moonshot released in July, through additional training for complex reasoning and tasks involving multiple steps and tools. Its predecessor comprises 896 neural networks but activates only 16 at a time to minimize inference costs.
Kimi K3 also uses Kimi Delta Attention, a custom implementation of Gated DeltaNet that reduces the RAM its attention mechanism needs to interpret prompts.
Moonshot says the upgraded model’s post-training consumed substantially more processing power than its earlier projects. The training system gained automated quality checks and infrastructure efficiency improvements. According to the company, there’s also a distinct Highspeed variant built to generate responses at two to three times the speed of the standard model.
Benchmark results split across coding, 3D work and finance
Moonshot’s reported results place Kim K3.1 Pro ahead of GPT-6 Astra in terminal-based programming and ahead of both rivals in financial analysis, but behind both in 3D model coding.
On Terminal-Bench 4.0, which tests agents on programming tasks, the model scored 65.7%, compared with Astra’s 58.2% and Opus 5.5’s 66.4%. Moonshot’s internal Kimi Code Bench 3.0 Long horizon score rose from its predecessor’s 55.95% to 61.2%.
The model scored 92.7% on BenchCAD, a 3D model coding benchmark, against Opus 5.5’s 96.2% and Astra’s 95.9%. It creates geometry through code and can animate the resulting models. Moonshot also reported improvements across Unity, Unreal Engine and Godot for generating interactive environments, websites and 3D assets.
In one demonstration, a walking simulator incorporated 296 skeletal components and 402 muscle components, with the model iterating through 167 development tasks and 420 reviews before achieving a character that walked continuously for 11.2 seconds. Another internal evaluation produced multiple 3D video games.
For financial work, Moonshot reported 64.4 on OfficeQA Pro v2, surpassing both rivals on an evaluation of financial document analysis, retrieval and numerical reasoning. Internal financial-modeling and financial-research results improved by 7.2 and 11.6 points, respectively, over Kimi K3.
Moonshot also noted that, based on its internal Kimi Hallu Bench Hard evaluation, hallucination rates dropped from 23.4% to 8.3%. Opus 5.5 recorded 15% with tools in the company’s comparison. Moonshot cautioned that this difficult-scenario test does not represent everyday hallucination rates.
Cybersecurity training changes and paid access
Moonshot deliberately excluded certain cybersecurity data and tasks from post-training to limit offensive capabilities while preserving general coding performance. The resulting model scored 81% on an internal vulnerability-research evaluation, versus 90.5% for a version trained without those restrictions.
Its cybersecurity benchmark performance trails Astra and Opus 5.5. Moonshot plans more advanced vulnerability-analysis features for future models.
Moonshot AI‘s Kim K3.1 Pro can be accessed via paid cloud offerings, with both the standard and Highspeed editions reachable through Kimi.ai, Kimi Work, Kimi Code, and the firm’s API. Moonshot plans to open-source the full model weights by November 4, allowing developers to run the model on their own infrastructure.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.