Spark X2.6 Turns One Prompt into a Working ESP32 Lab, a Fixed Volcano Bug, and a CSS Rubik's Cube
The gap between Chinese and Western frontier models just narrowed for agentic coding: a 293B MoE with a 1M context produces complete 3D apps and fixes physics bugs from a single prompt, and the recipes are public for anyone to verify.
Spark X2.6, iFlytek's new MoE model, is trained entirely on domestic Chinese compute and features 293B total parameters with 30B active per token, plus a 1M-token context window. The model is positioned for long-horizon tasks, code development, and Agent execution, and the tests here put that claim to work.
A round of prompts demanded zero external 3D assets and single-file delivery. They produced an ESP32 mini piano lab with React, TypeScript, and Three.js, complete with a three-step teaching interface and Arduino code; a volcano eruption demo where it diagnosed a hardcoded ground height (y=90) that made lava hover in mid-air and fixed it by computing terrain height from (x,z) coordinates and projecting velocity along the slope; a pure-CSS 3x3 Rubik's cube using integer coordinates and 3x3 orientation matrices; and a flight simulator with six-degree-of-freedom physics and procedural terrain.
All models are reported to have run to completion, though the flight simulator needed debugging. The full prompts are included for anyone to reproduce the tests, which is what makes this more than a marketing claim; it's a benchmark you can run yourself.
The Rubik's cube case is notable because it requires the model to maintain a consistent 3D state across 27 moving parts; the fact that it used 3x3 matrices rather than incrementally accumulating angles shows it reasoned about coordinate transforms.
The physics bug fix is a strong test: the model had to read the original code, identify the hardcoded constant, and apply geometric reasoning about slope normals, not just pattern-match a common error.
Marketing claims aside, the full prompts make the tests falsifiable: anyone can run them in their own Agent tool and compare.
Ever since DeepSeek open-sourced its reasoning model, all sorts of domestic models have started to pop up, haven't they? Right?