A Chinese lab just open-sourced a coding agent that's making developers stop paying for Claude Code.

Code Agent Kills Claude Code Kimi.ai Kimi-K2.7-Code, our latest coding model, is now released and open-sourced! Improved coding & agent performance over K2.6: +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite. Reasoning efficiency: Less overthinking, with 30% lower Python # manim -pqh gradient_descent_2d.py GradientDescent2D from manim import * import numpy as np class GradientDescent2D(Scene): def construct(self): # Objective and gradient def f(x, y): return (x - 1)**2 + (y + 1)**2 def grad(x, y): return np.array([2 * (x - 1), 2 * (y + 1)]) eta = 0.22 iters = 10 x = np.array([-4.0, 3.0]) # Axes plane = NumberPlane( x_range=[-5, 5, 1], y_range=[-4, 4, 1], background_line_style={"stroke_opacity": 0.3} ) x_label = plane.get_x_axis_label("x") y_label = plane.get_y_axis_label("y") title = Text("Gradient Descent in 2D", font_size=24) title.to_edge(UP) self.add(plane, x_label, y_label, title) # Contours levels = [0.5, 1, 2, 3, 5, 8, 12] contours = VGroup( *[ParametricFunction( lambda t: plane.coords_to_point(np.sqrt(c - (t-1)**2), t), t_range=[1-np.sqrt(c), 1+np.sqrt(c)], color=GRAY ) for c in levels] ) self.add(contours) # Path path = VMobject(color=YELLOW, stroke_width=2) path.set_points_as_corners([ plane.coords_to_point(x[0], x[1]), plane.coords_to_point(x[0] - eta * grad(x[0], x[1])[0], x[1] - eta * grad(x[0], x[1])[1]) ]) self.play(Create(path), run_time=1) # Animation dots = VGroup( *[Dot(plane.coords_to_point(x[0], x[1]), color=RED) for _ in range(iters)] ) self.play(FadeIn(dots[0])) self.add(dots) for i in range(iters - 1): x = x - eta * grad(x[0], x[1]) new_point = plane.coords_to_point(x[0], x[1]) path.add_points_as_corners([new_point]) dots[i+1].move_to(new_point) self.play(Create(path.points[i+1:]), FadeIn(dots[i+1]), run_time=0.5) self.wait(2) Gradient Descent in 2D (Contours + Path) f(x, y) = (x - 1)**2 + (y + 1)**2 n = 0.22 x = -4.00 y = 3.00 f(x,y) = 57.00 K2.7 AI MODEL LEADERBOARD INDEPENDENT BENCHMARK RESULTS K2.7 General benchmark suite 97.4 OPT-S General benchmark suite 95.1 Model C General benchmark suite 89.3 Model D General benchmark suite 86.7 Model E General benchmark suite 82.0 Pause K2.7 Leaderboard reveal · 16s loop KIMI Type "/" to invoke plugins and skills + K2.7 Instant Agent Swarm Slides Deep Research V build a build a landing page for my product, Meet SKIN DISCIPLIN Product Introduction For a fast way of life, the answer to health is actually a change of pace. The Purpose Contemporary SLOW DOWN Product Showcase and minutes 2- In- 1 Ingredients SLOW RELEASE FAST ACTING week test grades finds it itself bugs