Breaking

Moonshot AI's Kimi K3 Lands as 2.8-Trillion-Parameter Model Rivaling Claude and GPT at a Fraction of the Cost

July 17, 2026 at 21:45 EDT

Foundation Models
Software Dev & Coding
Open Source

Chinese startup Moonshot AI has released Kimi K3, a 2.8-trillion-parameter mixture-of-experts model that tops a leading front-end coding benchmark while undercutting rivals on price, positioning it as one of the most capable open-weight systems yet to emerge from China. The model, announced on July 16, 2026, has drawn attention for reconstructing complex interfaces from minimal input, including a 3D globe dashboard rebuilt from a single prompt and one reference image with textures, lighting and glass panels rendered close to pixel-perfect.

July 16, 2026 · Moonshot AI

Kimi K3: China's 2.8-Trillion-Parameter Open Coder Tops the Charts — at a Third of the Price

A sparse mixture-of-experts model with native vision and a one-million-token context takes first place on a leading front-end coding benchmark, positioning it as one of the most capable open-weight systems yet from China.

2.8T

total parameters (MoE) — first ~3T-class model shipped with accessible weights

token context window, native vision, tuned for long-horizon agentic coding

1.8%

experts active per token — just 16 of 896 fire at once

Frontend Code Arena — Benchmark Score

Kimi K3 takes first place, edging out Claude Fable 5.

1,679

Kimi K3

1,631

Claude Fable 5

Standout areas: brand & marketing · reference-based design · data & analytics · simulations

Output Price — per Million Tokens

Roughly a third of the cost of Claude Fable 5 — "Sonnet prices."

$15

Kimi K3

~$45

Fable 5 (approx.)

Input $3.00 (cache miss) → $0.30 (cache hit) · flat pricing across the full context window

The Enthusiasm

Developers report generating intricate 3D interfaces, interactive sites and data visualizations in a single shot — including a globe dashboard rebuilt near pixel-perfect from one image. Several say it beats Fable 5 and GPT-5.6 on graphics- and game-adjacent tasks.

The Caveats

Weights are not yet public, so independent verification is limited — current evidence rests on vendor benchmarks and demos. The consumer app also caps monthly usage, and even light sessions can draw down quota quickly.

Roadmap to the Weight Drop

Jul 16, 2026

Model announced

API live immediately

→

Jul 27, 2026

Weights released

Modified MIT license

→

Coming weeks

Independent scrutiny

Do the leads hold up?

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.