Kimi K3.1 Pro Released: Building Complete Games, Editable 3D Models, and Linked Financial Spreadsheets

Moonshot AI released Kimi K3.1 Pro, the first model in its K3.1 family, alongside Kimi K3.1 Pro Highspeed, which the company says delivers 2x to 3x faster response times. The models are available across the Kimi web, desktop, and mobile apps as well as the API, while the full model weights await subsequent public release.

In the officially disclosed specifications, 2.8T total parameters equal roughly 2.8 trillion, with 104B active parameters representing approximately 104 billion. The model uses a Mixture-of-Experts (MoE) architecture where each calculation activates only the subset of expert networks relevant to the current task—meaning the overall model is not limited to 104B—and supports multimodal input. The 1 million token limit represents the maximum context window per request, which does not translate to 1 million Chinese characters due to tokenization rules.

This release introduces several key shifts:

• Connecting fragmented tools into complete projects: In game engines like Unity, Unreal Engine, or Godot, chaining web research, 3D asset generation, scene assembly, and gameplay scripting to deliver runnable projects.
• Parametrically rebuilding 3D models from numeric edits: Writing and executing Blender and CAD scripts to update component dimensions and scene structures by adjusting parameters.
• Cross-verifying financial and fiscal data across tables: Directly reading chart-dense materials and linking multiple financial statements with traceable formulas, while reducing errors in internal factual stress tests compared to the previous generation.
• Independently diagnosing deep-seated bugs in simulation environments: Leveraging post-training reinforcement learning to run continuous tests, diagnose anomalies, and locate underlying logical errors during complex, long-horizon tasks.

From Writing Code to Running Complete Projects

Shipping a complete software or game project requires more than writing isolated snippets of code; it demands integrating art assets, business logic, scene assembly, and runtime testing into a cohesive whole. For K3.1 Pro, the technical roadmap substantially scaled up post-training compute relative to K3, conducting joint reinforcement learning across hundreds of downstream domains. In practical terms, this trains the model to operate tools directly inside target task environments and continuously refine its methods based on execution feedback. To support this, the engineering team optimized underlying infrastructure for long-context reasoning, sandbox isolation, task scheduling, and environment health checks.

An official Godot 4.7.2 action-adventure game demo brings these capabilities together: five explorable zones feature terrain, roads, water bodies, vegetation, and villages. Characters can engage in third-person combat, fight a two-stage boss, gather resources, experience day-night cycles, and fast-travel between campfires. The project incorporates ready-made asset packs and fonts from Quaternius, Kenney, and KayKit, with the model tasked with tying together external assets, procedural scenes, and gameplay logic into a playable game.

Long-horizon tasks also require the model to uncover root causes behind anomalous results. In a musculoskeletal human walking simulation disclosed in the official technical blog, the model built a gait controller based on the open-source anatomical model human-atlas (comprising 296 skeletal components including cartilage, alongside 402 muscle components). Over a long-horizon cycle consisting of 167 subtasks and 420 independent reviews, the simulated human progressed from losing balance and collapsing within seconds to walking continuously with alternating strides for 11.2 seconds.

When adjusting reward shaping—the scoring rules defining what makes a good walk—failed to produce natural arm swings, the model did not simply keep tweaking those weights. Instead, it independently tested the physics engine's simulation dynamics, examined the formulas governing how shoulder muscle forces rotate joints, discovered and corrected 36 moment-arm calculation errors, and only then resumed controller training.

This capacity for iterative debugging across command lines and sandbox environments is also reflected in benchmark results and operational boundaries. On Terminal-Bench 4.0, a strictly constrained real-world terminal debugging benchmark, K3.1 Pro scored 65.7%, close to Claude Opus 5.5 (66.4%) and ahead of GPT-6 Astra (58.2%). On the long-horizon software engineering benchmark SWE-Marathon, it reached 51.7%, topping Claude Opus 5.5 (46.3%). On the internal long-horizon coding benchmark Kimi Code Bench 3.0 Long-horizon, K3.1 Pro scored 61.2%, up from the previous K3's 55.95%, but behind GPT-6 Astra's 66.5%.

On the scientific terminal benchmark Terminal-Bench Science 0.1, K3.1 Pro recorded 38.1%, trailing GPT-6 Astra (68.1%) and Claude Opus 5.5 (63.3%), highlighting its limitations in specialized scientific terminal tasks.

In tool-assisted knowledge and system-operation benchmarks, K3.1 Pro reached 70.0% on HLE with tools (higher than Opus 5.5's 67.7%) and 79.6% on OSWorld (slightly below Opus 5.5's 81.8%).

These scores were compiled and reported by Kimi. Most models were evaluated at their highest reasoning effort settings, but the evaluation harnesses were not entirely uniform, and certain comparative scores were drawn from published reports or peak leaderboard entries. These figures should be evaluated alongside their corresponding tool environments.

Why Generated 3D Models Can Still Be Resized

In 3D modeling, conventional generative models often output static polygonal meshes. While these meshes deliver complete visual geometry, modifying structural dimensions typically requires manually editing vertices or regenerating the entire model from scratch.

Parametric modeling differs by generating programmatic scripts that construct geometry procedurally. If you are building a mechanical part with mounting holes, the script defines its length, width, height, and hole diameters as explicit numerical variables. When hole dimensions need adjusting later, modifying the corresponding numeric values and re-running the script instructs the software to rebuild the model under the new parameters. This approach preserves ongoing editability through parameters, allowing updates to regenerate along the original modeling pipeline.

K3.1 Pro can write and execute these scripts for Blender and CAD software. The claim of requiring 'no external MCP connectors' means it drives the software through script execution; operating Blender or CAD locally still requires the corresponding software and runtime environment installed on the host machine.

Scripted modeling can also embed part hierarchies, assembly sequences, and kinematic relationships into the model, feeding 3D outputs directly into downstream editing, teardown, or 3D printing pipelines.

For instance, an official turbofan engine model was scripted in FreeCAD using Python, containing 2,100 solid entities across the fan, gearbox, compressor, combustor, turbine, and external piping. The video demonstrates sequential assembly of these components before revealing internal structures through transparent casings and exploded views; the fan, low-pressure shaft, and high-pressure shaft rotate at 2,780, 8,590, and 12,400 RPM respectively. While it showcases component assembly and kinematic relationships, whether it meets real-world manufacturing standards requires separate engineering validation.

Another modular lamp model built with Trimesh connects 3D modeling to physical fabrication: the circular silhouette of Fujian Tulou architecture is transformed into an openwork lampshade split into five 3D-printable parts. The official video presents component breakdown and assembly before capturing 3D printing in progress, the finished white lampshade, and illumination testing, illustrating how digital assets transition into physical fabrication and assembly.

Beyond executing modeling scripts offline, the model can deliver interactive 3D experiences that run directly in web browsers without specialized client software:

• 3D Car Showcase: Built with custom Three.js, allowing users to rotate camera angles, click to inspect specific mechanical components, drag sliders to trigger exploded-view disassembly and reassembly, and toggle paint finishes, with cloth unveilings and camera moves baked in Blender;
• 3D Robotic Arm Showcase: Supports orbit viewing, toggling gestures and viewpoints, making the shell transparent or hiding it entirely, and exploding internal components for inspection before reassembly.

In the BenchCAD evaluation—which reconstructs CAD code from multi-view renderings—the team randomly sampled 1,000 cases from the 17,900-case Vision2Code dataset. The benchmark measures Voxel Intersection over Union (Voxel IoU, %) to assess spatial volume overlap between the generated 3D geometry and the ground-truth reference model. This reflects overall geometric shape alignment, not manufacturing tolerances or assembly fit in precision industrial engineering.

Under benchmark conditions, K3.1 Pro with tool use achieved an average Voxel IoU of 92.7%, approaching Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%). In single-pass generation tests without external tools, K3.1 Pro scored 72.8%, on par with Claude Opus 5.5's 73.0%.

On the product side, the Kimi platform simultaneously introduced 3D Make, allowing users to enter prompts online, preview and edit generated 3D models, download 3D-printable files directly, or export videos and Live Photos for sharing.

Financial Spreadsheets Require Interlinked Formulas, Research Demands Traceable Sources

In financial modeling, analysts place a premium on dynamic consistency across data. In the official Nvidia (NVDA) model demo, 'formula-linked' means that adjusting revenue growth rates or cost of goods sold assumptions in the workbook triggers automatic updates across financial statements and valuation calculations, enabling anyone to trace the provenance of every figure backwards through cell formulas. While the main announcement post mentioned discounted cash flow (DCF) models that discount future cash flows to present value and leveraged buyout (LBO) models that evaluate debt-financed returns and debt service, the downloadable NVDA workbook from the technical blog is a 16-sheet valuation model centered on DCF, with no standalone LBO sheet. In addition, the company showcased an investment committee deck for CATL and a quantum computing industry research report with multi-tiered citations.

The original files are available for all three examples, allowing readers to trace their calculations and evidence:

• Nvidia Financial Model (Download XLSX): Examine operating assumptions in Assumptions, trace formulas across the income statement, balance sheet, and cash flow statement (IS, BS, CF), inspect balance checks in Checks, and verify valuations in DCF and Valuation alongside citations in Data_Sources. This download does not include an independent LBO-named sheet.
• CATL Investment Committee Deck (Download PPTX / View PDF): Outlines consensus expectations, reconciles variances against reported financials and company guidance, and flags subsequent events and downside risks. Sources are stratified across company filings, databases, broker reports, and proprietary calculations, with conflicting figures displayed side by side.
• Quantum Computing Sector Report (View PDF): Analyzes technical roadmaps, performance thresholds, and market expectations, linking assertions back to supporting evidence and follow-up verification milestones, backed by hundreds of end-of-text citations. Citations span multiple source tiers and explicitly label unverified working drafts, enabling readers to audit specific claims.

On OfficeQA Pro v2, a benchmark evaluating public finance documents and fiscal data analysis, K3.1 Pro scored 64.4 points. The evaluation feeds full rendered PDF pages without extractable machine text, testing cross-page visual grounding and multi-step numerical calculation. Moonshot AI said it assigned zero scores in this evaluation to certain external-model runs that searched the web for reference answers. On Frontier Finance, which evaluates end-to-end investment workflows, K3.1 Pro scored 58.6, behind Claude Fable 5.1 (60.8 points).

Across three internal finance benchmarks disclosed by Moonshot AI, metrics showed gains across the board: Financial Modeling (assessing formula linking and workbook integrity) rose from 50.4 to 57.6 (+7.2 points); Financial Research (evaluating evidentiary auditability and temporal cutoff rigor) advanced from 58.8 to 70.4 (+11.6 points); and Business Operation (testing multi-step operational decision-making in simulated environments) jumped from 8.9 to 37.0 (roughly 4.2x the prior generation; this represents simulated task performance, not actual commercial returns).

The officially reported '65% hallucination reduction' measures the error rate among verifiable factual claims in responses. When the model produces an analysis, its output is broken down into atomic assertions (such as a company's founding year, quarterly revenue figures, or technology release dates), which a fact-checking Agent verifies individually.

In the internal high-difficulty stress test (Kimi Hallu Bench Hard), the previous-generation K3 exhibited a claim-level factual error rate of 23.4%, which dropped to 8.3% for K3.1 Pro. This represents a relative reduction of (23.4% - 8.3%) ÷ 23.4% ≈ 64.53% (around 65%), or an absolute drop of 15.1 percentage points. Meanwhile, the proportion of responses containing at least one factual error fell to 42.4% of the prior generation's rate—indicating that flawed responses occur far less frequently than before, not that 42.4% of responses are currently erroneous. Because this test specifically selects confusing boundary queries to stress-test models, these figures do not represent error rates in everyday conversational usage. Furthermore, under identical Hard conditions, GPT-6 Astra achieved a claim-level error rate of 5.5%, remaining lower than K3.1 Pro's 8.3%.

Safety Trade-offs Ahead of Open Weights

As post-training reinforcement learning scales up, generalized proficiency in long-horizon code debugging naturally extends into systems security and exploit development. To mitigate potential misuse risks before releasing open weights, Moonshot AI deliberately reduced certain capabilities during model training.

While preparing mid-training data and reinforcement learning environments, the team identified and removed data directly related to cybersecurity exploitation, while retaining general-purpose coding tasks essential for everyday programming capabilities.

According to evaluations on the internal benchmark xdev-bench-hard (conducted without hard timeouts or output token limits), this targeted filtering yielded clear differences:

• Without filtering cyberattack-related data, the model achieved a pass@3 success rate (solving the task in at least one of three attempts) of 90.5%;
• Following targeted data removal, the shipping K3.1 Pro's pass@3 rate fell back to 81.0%;
• Despite intentional risk mitigation, K3.1 Pro still scored substantially higher than the prior K3's 33.3%, indicating that generalized debugging experience on complex code still transfers to vulnerability workflows;
• While the closed-source model GPT-5.6-Sol achieved 100% pass@3 in the same chart, it originated from a third party and ran under a different harness, precluding direct head-to-head ranking under identical conditions.

Moonshot AI framed this as a deliberate safety consideration for open weights: aiming to avoid expanding frontier offensive capabilities, while distributing auditing and debugging tools to a broader developer base to inspect unmaintained legacy software. The full weights are not yet publicly available.

Access Channels and End-to-End Task Costs

Users can access K3.1 Pro through the following channels:

1. Multi-platform interfaces: Visit the Kimi web client, download or update mobile apps (supporting iOS, Android, and HarmonyOS), or use Kimi Work on desktop (version 3.1.0 or later, supporting Windows and Apple Silicon Mac).
2. Engineering and coding: Inside Kimi Code, run the /model command and select the display name Kimi K3.1 Pro.
3. Enterprise deployment: Kimi Enterprise administrators can enable access on demand via the admin console; it is disabled by default upon launch.
4. Model weights: The company committed to making the full weights publicly available by November 4 at the latest.

For developer access, the official API platform offers two model IDs:

• Standard Edition (kimi-k3.1-pro): Priced at $2.00 / MTok (per million tokens) for input, $10.00 / MTok for output, and $0.20 / MTok for cache-hit input. The company notes this base pricing is roughly two-thirds of K3.
• Highspeed Edition (kimi-k3.1-pro-highspeed): Priced at $4.00 / MTok for input and $20.00 / MTok for output—double the standard baseline—with cache hits also at $0.20 / MTok. While advertised as 2x to 3x faster, end-to-end projects still involve tool execution and file I/O overhead, so overall task turnaround cannot be guaranteed to scale linearly. In addition, Zero Data Retention is available as an optional privacy feature for API clients.

When budgeting for long-horizon Agent workloads, developers must distinguish between per-call rates and cumulative multi-turn expenditures. In complex end-to-end engineering, an Agent continuously ingests logs, executes tests, and iterates through self-corrections, compounding context size with every step.

Taking statistics from the internal xdev-bench-hard benchmark as an example, completing an average successful trajectory required a cumulative input context of 159.8 million tokens (159.8 MTok) and cumulative output of 546 thousand tokens (546 KTok). This represents total input accumulated across multiple calls over the entire task; individual calls remain subject to the 1M context window limit, and this does not reflect consumption for everyday end users. At Standard edition rates, assuming zero context caching for illustrative calculation, estimated token charges for a single successful run would be $325.06 (159.8 × $2.00 + 0.546 × $10.00). This is purely a multiplication example rather than an actual billed amount, and omits costs from failed attempts or environment setup.

Prompt Cache mitigates these costs by billing repeated historical prefixes at $0.20 / MTok when matched. Readers can adjust assumed cache hit rates in the calculator above; actual cost reductions depend entirely on realized cache hit rates, with no guaranteed fixed percentage savings.

Additionally, an official Terminal-Bench 4.0 task cost analysis indicates that, under the specific configuration tested, K3.1 Pro incurred lower average costs per problem while achieving comparable terminal debugging scores. However, this finding is rooted in specific benchmark harnesses and pricing structures that differ from production environments. Furthermore, benchmark criteria for comparator models in this scatter plot diverge from the main overview charts, meaning they cannot be synthesized into a universal price-performance ratio.