TECH

Moonshot AI Launches New Model With Improved Coding and Agent Capabilities

Published: Apr. 21, 2026 5:54 p.m. GMT+8

00:00

00:00/00:00

Listen to this article 1x

Chinese artificial intelligence startup Moonshot AI has launched an open-source large model that it claims rivals top foreign competitors like OpenAI’s GPT-5.4 in coding and agent-based tasks.

Released on Monday, the Kimi K2.6 model is designed to execute extended programming assignments and coordinate a large cluster of AI agents for complex output. The new model is available across the company’s platforms, including its website, app and Kimi Code assistant.

You've accessed an article available only to subscribers

Subscribe today for just $.99.

VIEW OPTIONS

Unlock exclusive discounts with a Caixin group subscription — ideal for teams and organizations.

Subscribe to Save an extra $50. Introductory offer for new readers. Subscribe now.

Share this article

Open WeChat and scan the QR code

DIGEST HUB

Digest Hub

Back

Explore the story in 30 seconds

Moonshot AI launched open-source Kimi K2.6, rivaling GPT-5.4 in coding and agent tasks.
Matches/outperforms GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro on SWE-Bench Pro and DeepSearchQA.
Codes 4,000+ lines over 13 hours; coordinates 300 sub-agents for complex outputs; praised on Reddit.

AI generated, for reference only

Who’s Who

Moonshot AI: Moonshot AI, a Chinese startup, launched the open-source Kimi K2.6 model, rivaling GPT-5.4 in coding and agent tasks. It excels on SWE-Bench Pro and DeepSearchQA benchmarks, handles 13-hour sessions with 4,000+ code lines, and coordinates 300 sub-agents for complex workflows like building websites and presentations. Available on its platforms, it's praised on Reddit for front-end development.

OpenAI: Moonshot AI claims its open-source Kimi K2.6 model rivals OpenAI's GPT-5.4 in coding and agent-based tasks, matching or outperforming it on benchmarks like SWE-Bench Pro and DeepSearchQA.

Anthropic: Moonshot AI's K2.6 model matched or outperformed Anthropic’s Claude Opus 4.6 on benchmarks like SWE-Bench Pro (software engineering) and DeepSearchQA (AI agent research).

Google: Moonshot AI's K2.6 model matched or outperformed Google’s Gemini 3.1 Pro on benchmarks like SWE-Bench Pro (software engineering) and DeepSearchQA (AI agents’ deep web research).

AI generated, for reference only