Artificial Intelligence

Kimi K3.1 Pro Drops Overnight: 2.8T Parameters, Terminal Score Just 0.7 Behind Opus

By 老金带你玩AI · Oct 24, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A Chinese lab closed the gap on terminal-based agentic coding and CAD generation, with weights promised open. Western teams evaluating model contracts need a concrete alternative that matches Opus on tool-heavy workflows and beats it on price transparency, assuming the open release lands as scheduled.

Summary

Kimi K3.1 Pro launched overnight. The model packs 2.8 trillion total parameters and a 1 million token context window. Official demos show it producing complete 3D adventure games in Godot, an Unreal Engine 5.8 cinematic with a 36-million-triangle dragon scene, and a Blender-rendered robot short film. The most tangible artifact: a lamp designed from Fujian tulou inspiration, printed on a 3D printer, assembled, and lit.

Benchmarks place the model close to top-tier rivals. On Terminal-Bench 4.0, K3.1 Pro scores 65.7% against Claude Opus 5.5's 66.4% and GPT-6 Astra's 58.2%. On BenchCAD with tools, it reaches 92.7%, behind Opus (96.2%) and Astra (95.9%). An internal long-horizon coding test shows 61.2%, up 5.25 points from the previous K3. The full weight release is scheduled by November 4.

The demo also includes an agentic task: a walking simulator that initially failed, then the model traced 36 calculation errors in shoulder muscle physics, fixed them, and achieved a continuous 11.2-second walk. Office outputs include an interconnected Nvidia financial model with live formulas and a report with traceable citations.

Takeaways

— Kimi K3.1 Pro has 2.8 trillion total parameters and a 1-million-token context window.

— Complete model weights are planned for release before November 4.

— Terminal-Bench 4.0 score is 65.7%, versus Claude Opus 5.5 at 66.4% and GPT-6 Astra at 58.2%.

— BenchCAD with tools score is 92.7%; Opus 96.2%, Astra 95.9%.

— Internal long-horizon coding test improved from 55.95% to 61.2%.

— Demos included a full Godot adventure game with 5 regions and a two-phase boss.

— A 3D-printed lamp model was designed in 5 parts, printed, assembled, and lit.

— A walking simulator exposed 36 errors in shoulder muscle calculations; fixing them produced an 11.2-second continuous walk.

— Sendable finance models had linked spreadsheets with formulas and traceable citations in reports.

Conclusions

The 36-error debugging story shows the model's agentic loop not just executing code but inspecting underlying simulation code and fixing mathematical physics, which raises the bar for what tool-using AI can self-correct.

Pairing a 1M context window with open weights before November 4 puts pressure on closed labs to justify their API pricing, especially for teams that need long-context engineering tasks.

The benchmark gap to Opus on terminal tasks is small enough that workflow differences (tooling, prompting, evaluation harness) may matter more than raw model capability.

Concepts & terms

Token

A basic unit of text that an AI model reads or writes. Roughly equivalent to a word fragment or a small chunk of characters; 1 million tokens allow processing a very long document or conversation in one go.

Parametric modeling

A CAD approach where model dimensions are driven by parameters. Change a width, and dependent parts like shelves and openings recalculate automatically, keeping the design consistent.

Terminal-Bench

A benchmark that tests an AI's ability to complete real engineering tasks in a command-line terminal, including running commands, fixing errors, and navigating filesystem operations.

BenchCAD

A benchmark that gives an AI multiple images of a 3D object and asks it to write CAD code to recreate the model, optionally with tool use.

Long-horizon task

A multi-step task that requires an AI to plan and execute many operations sequentially, often involving tool calls, without losing track of the overall goal.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗