Breaking

Moonshot AI Launches Kimi K3, a 2.8-Trillion-Parameter Model, With Open Weights Due July 27

July 16, 2026 at 15:30 EDT

Foundation Models
Open Source
AI Agents

Moonshot AI has launched Kimi K3, a new flagship model the Beijing-based lab describes as its first open-source system at the 2.8-trillion-parameter scale, with benchmark results it says rival proprietary models from leading U.S. labs. The model went live on July 16 across the Kimi app, Kimi Code and the company's API, while the open weights are slated to arrive on July 27.

July 16 · Moonshot AI

Kimi K3: A 2.8-Trillion-Parameter Open Challenger

Moonshot's new flagship claims benchmark parity with leading U.S. proprietary models — moving to an estimated 2.8T parameters and a ~1M-token context. Open weights land July 27, when independent researchers can finally test the claims.

2.8T

Estimated total parameters (K2 was 1T)

~1M

Context window in tokens (K2 was 256K)

91.2

Single-agent BrowseComp, reported state-of-the-art

Generation-over-generation scale

Total parameter count: K2 vs K3 — roughly a 2.8× jump.

2.8T

Where K3 lands on the leaderboards

Reported scores against leading proprietary models (higher is better).

GDPval-AA v2

Kimi K3

1687

Opus 4.8 Max

1600

AA-Briefcase

Kimi K3

1527

GPT-5.6 Sol Max

1495

Early praise

Strong on complex, long-horizon workflows
Improved agentic performance in production-style tests
Competitive 3D & coding output vs Fable 5

Open questions

Slow responses — some runs over an hour
Heavy token consumption raises efficiency doubts
Strengths skew toward front-end & coding

API Pricing · per million tokens

Input

$15

Output

$0.30

Cached input

The decisive test is July 27 — when open weights ship and independent researchers can validate the benchmark claims and probe real-world performance.

Continue reading

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