loading
loading
Moonshot's Kimi K3.1 Pro Challenges Claude With 2.8 Trillion Parameters Moonshot's 2.8T parameter MoE now runs full Unreal projects, parametric CAD, and finance modeling with 65% fewer hallucinations. - Moonshot released Kimi K3.1 Pro, a 2.8T total / 104B active MoE with 1M-token context. - Scores 65.7% on Terminal-Bench 4.0, between GPT-6 Astra (58.2%) and Claude Opus 5.5 (66.4%). - Hits 92.7% on BenchCAD with tools, approaching Opus 5.5 and GPT-6 Astra. - Finance hallucinations down 65%; state-of-the-art 64.4 on OfficeQA Pro v2. - Live on platform.kimi.ai, Kimi Work, and Kimi Code; Highspeed variant is 2-3x faster. - Open weights promised within roughly two weeks of launch. Moonshot has launched Kimi K3.1 Pro, a large mixture-of-experts model built for long-running agent workloads. Its focus includes multi-file software development, parametric 3D modeling, and financial analysis, areas often served by frontier closed models. The hosted model is available through Kimi Work, Kimi Code, and the API, while downloadable weights are scheduled to follow. Three trillion parameters, 104 billion active K3.1 Pro activates a fraction of its parameters for each token, reducing computation relative to a dense 2.8-trillion-parameter model. The complete set of expert weights still requires substantial storage and distributed infrastructure. Its million-token context window is designed for large repositories, filings, transcripts, and other document collections. Moonshot released the original Kimi K3 on 16 July 2026 with the same 2.8-trillion-parameter scale. K3.1 Pro retains that architectural foundation while emphasizing agentic tool use, where the model operates shells, editors, tests, and domain applications across multiple steps. Terminal work gets the headline score Terminal-Bench 4.0 evaluates agents on resource-constrained command-line tasks with task-specific tests. An agent must inspect repositories, edit files, execute commands, diagnose failures, and complete the requested work inside a sandboxed environment. Harness choice can materially change an agent benchmark score because the surrounding software controls prompts, tool calls, retries, context management, and command execution. Earlier Kimi evaluations showed gaps between Moonshot’s Kimi Code harness and independent runs. The 65.7% result therefore remains a vendor-reported figure pending replication under a shared harness and resource budget. Moonshot also published results for XDev-Bench Hard, another coding-agent evaluation. The chart below reflects the company’s reported comparison and carries the same need for independent reproduction. 3D agents span the production stack Moonshot has tuned K3.1 Pro to operate Unreal Engine, Unity, Godot, Blender, and CAD software. On BenchCAD, which evaluates 3D reconstruction from multi-view renderings through generated CAD code, the model scores 92.7% in a tool-enabled configuration. Moonshot reports 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra under its comparison. The intended workflow extends from asset construction through scene assembly, animation, and rendering. That scope gives game, visualization, and CAD teams a concrete evaluation target: whether one agent can preserve project state and geometry constraints while moving among code, assets, and application-specific tools. Finance gains need workload-level checks Moonshot reports a score of 64.4 on OfficeQA Pro v2 and describes it as state of the art. Its internal evaluations show gains over K3 of 7.2 points in financial modeling and 11.6 points in financial research, along with roughly 4.2 times K3’s score in simulated business operations. The internal results lack independent replication. Demonstrations include an Nvidia discounted cash flow and leveraged buyout model with linked statements and traceable formulas, a CATL investment committee deck, and a quantum-computing sector report with source-level citations. These examples define the target workflows; production evaluations still need to measure formula accuracy, citation fidelity, spreadsheet integrity, and source coverage. Moonshot also claims a 65% reduction in hallucinations during finance work. The million-token context window can accommodate large filings and transcript collections, while reliability depends on retrieval quality and arithmetic consistency across long runs. Teams evaluating the model can test whether every quoted figure resolves to a source and whether spreadsheet changes propagate correctly through linked cells. Cloud access now, weights later At launch, users can access the model through Kimi Work for document and spreadsheet tasks, Kimi Code for command-line software development, and the Kimi API. Moonshot says open weights will arrive on an approximately two-week schedule, so self-hosting depends on a later release. Open weights bring hardware and license costs The original K3 weight repository occupies roughly 1.56 terabytes, and K3.1 Pro belongs to the same architecture class. A production deployment will require multiple GPUs and may require multiple nodes, depending on precision, quantization, throughput targets, and available accelerator memory. Runtime capacity must also cover the key-value cache, framework overhead, and long-context requests. Commercial terms require separate review. The custom K3 license imposed revenue sharing of up to 30% on inference providers generating more than US$20 million annually. Moonshot has not confirmed whether K3.1 Pro will inherit those terms. Hosted inference providers need the final K3.1 Pro license before setting pricing or committing infrastructure. Benchmarks to run before switching - Reproduce coding tasks with the same harness, retry policy, tool permissions, time limits, and token budget used for the incumbent model. - Test repository-scale changes that require navigation, implementation, test execution, and recovery from failed attempts. - Measure 3D project validity inside the target engine or CAD package, including geometry constraints, asset references, and render completion. - Audit financial outputs at the cell, formula, citation, and source-document levels. - Compare K3.1 Pro and Highspeed on latency, throughput, tool-call reliability, and total cost per completed task. - Evaluate long-context retrieval using documents whose correct answers and source locations are known. - Confirm weight availability, serving requirements, and commercial license terms before planning a self-hosted deployment. K3.1 Pro offers a concrete open-weight candidate for coding, 3D, and finance agents if Moonshot’s results hold under independent testing. Its practical value will depend on harness-controlled evaluations, serving economics, and the final license attached to the downloadable weights.
document fetched 2026-10-24 15:22