Open-Source LLM Sets New Benchmark: A Deep Dive into Code Intelligence and 3D Creation Capabilities

Model Capability Upgrade Overview

The open-source large language model landscape has received a significant version update. Building on its previously released base model, this release applies extensive post-training optimization to achieve substantial gains across multiple core capability dimensions. According to official benchmark results, this update is not a simple parameter scaling exercise—it represents a substantive capability重构 (reconstruction) in critical scenarios including code execution, spatial reasoning, and professional document understanding. Its overall performance now positions it as a viable competitor to leading closed-source models globally.

Notably, this iteration pushes performance to new heights while preserving its open-source nature. For technical teams, this means open-source solutions are no longer a “compromise choice” on the path toward model autonomy and cost optimization—they are increasingly becoming strategic options capable of supporting core business requirements.

Code Capabilities: From Completion to Autonomous Execution

Breakthrough Progress on Terminal Bench 4.0

On the Terminal Bench 4.0 benchmark—which measures a model’s real-world coding ability—this model achieved a score of 65.7. This score not only makes it the first open-source model to enter the top tier on this benchmark, but also places it above certain leading international closed-source products in the rankings. Terminal Bench is regarded as a touchstone because it simulates a developer’s actual terminal environment, requiring the model not only to understand code logic but also to execute commands, parse outputs, and handle exceptions. It is a hardcore indicator of whether a model can truly “get work done.”

The deeper significance of this result is that open-source models have moved from “assisted completion” to “autonomous execution” in the code domain. The model can understand multi-step engineering tasks and progressively complete operations within a terminal environment, opening up possibilities for automated operations, intelligent CI/CD, and similar scenarios.

An Optimal Balance of Cost Efficiency

Beyond performance, the model is equally notable for its cost control. According to official statistics, it is one of the most cost-efficient frontier models on Terminal Bench 4.0 tasks per unit cost. While lighter-weight models with lower unit prices exist on the market, their task completion rates fall far short of this version. This comparison clearly reveals a trend: in professional scenarios, low price alone is no longer decisive—the performance-to-cost ratio is the core criterion for enterprise model selection.

For small and mid-sized teams with limited budgets, the arrival of high-performance open-source models means they can obtain code service capabilities approaching those of top closed-source models through private deployment, without incurring high API call fees. This helps narrow the technical gap between teams of different sizes to a certain extent.

3D Creation: A New Paradigm for Spatial Intelligence

Direct Scripting for Professional Modeling Tools

In the realm of 3D modeling and spatial reasoning, this update delivers highly disruptive capability enhancements. On CAD benchmarks that rely on external tools, the model scored 92.7—trailing leading international closed-source models by only a narrow margin and setting the best record for an open-source model on this evaluation.

More exciting is the breakthrough in engineering practice: the model can now directly write scripts for professional modeling tools such as Blender and CAD, completing the construction, refinement, and precise adjustment of layout, materials, and geometric properties of complex assets without relying on an intermediate interpretation layer. This “natural language to modeling tool” capability significantly lowers the barrier to entry for 3D creation.

According to officially showcased examples—whether constructing an indoor airport forest scene with a complex vegetation system, designing a vortex engine model with precise fluid dynamics structures, or generating a lamp model ready for direct printing—the model can output rigorously structured, executable script code. This means designers can devote more energy to creative ideation while delegating tedious parameter tuning to AI.

Online Interaction and a Closed-Loop Ecosystem

Alongside the model capability upgrade, the official platform has launched an online 3D creation feature—Kimi 3D Make. This feature supports online preview and editing of models; completed designs can be downloaded directly for 3D printing or exported as videos or animated images for sharing. This productization effort connects the full chain of “creative generation—model editing—physical output,” bringing 3D content creation from professional studios to a broader creative audience.

For practitioners in industrial design, game art, film and television visual effects, and related fields, this capability means efficiency during the prototype validation stage will increase exponentially. Work that previously required days of modeling cycles may now be completed with just a few hours of iterative optimization.

Finance and Document Intelligence: Depth in Vertical Domains

Leading Advantage in Complex Document Analysis

On the OfficeQA Pro v2 benchmark, which targets real-world financial scenarios, the model scored 64.4, ahead of all compared models. This test focuses on evaluating a model’s ability to perform complex data analysis based on real financial documents, covering high-difficulty tasks such as information retrieval, cross-document verification, and multi-step numerical reasoning.

Real financial documents are often complex in format, dense in terminology, and intertwined in numerical relationships, placing extremely high demands on a model’s semantic understanding and reasoning capabilities. The model’s leading performance demonstrates solid proficiency in long-text processing and structured information extraction, enabling it to address practical business needs such as investment bank research report analysis, listed company financial statement comparison, and risk and compliance review.

Investment Workflow Support

On the Frontier Finance benchmark, which simulates a complete investment decision-making process, the model scored 58.6—a significant improvement over the previous version. This test covers the full workflow from investment target screening and industry research to financial modeling and earnings forecasting, requiring the model to possess cross-domain knowledge integration and logical coherence.

This capability improvement positions the model as a potentially efficient assistant for financial analysts—not only can it quickly organize massive amounts of information, but it can also help generate structured draft research reports under a given analytical framework, allowing analysts to focus their core energy on in-depth judgment and strategy formulation.

Effective Hallucination Control

On the closely watched hallucination tests, the model also delivered satisfactory results. According to internal stress test data, its error rate dropped significantly from 23.4% in the previous version to 8.3%, trailing only top closed-source models globally. Hallucination has long been one of the core obstacles to deploying large models in serious scenarios, and the reduction in error rates directly affects the model’s usability in high-compliance domains such as finance, healthcare, and law.

Safety Capabilities and Future Outlook

In terms of cybersecurity capabilities, the official team observed that as post-training scale continues to expand, the model’s vulnerability research capabilities have shown a steady upward trend. This means the model can not only “understand” security knowledge but also assist in code auditing and vulnerability discovery in practical scenarios, providing strong support for security engineers.

Overall, this open-source model version update represents a milestone leap across three core domains: code intelligence, 3D creation, and professional document processing. It sends a clear signal to the industry: high performance and high cost-effectiveness are not mutually exclusive. As the open-source ecosystem continues to flourish, we have reason to expect that future large model applications will no longer be an exclusive game for a few giants, but rather infrastructure for the digital transformation of every industry. For technical decision-makers, now is the time to re-examine the strategic value of open-source models in enterprise intelligentization planning.