Loading video...
Video Failed to Load
Kimi K2.5 has arrived! 🥝 Here are 2 things to know: Aesthetic Coding x Agent Swarm.
630,014 views • 9 months ago •via X (Twitter)
35 Comments

👀

huge congrats 🔥

first clawdbot now kimi no excuses

Agent Swarm looks interesting. Trained with Parallel-Agent Reinforcement Learning (PARL), K2.5 learns to self-direct an agent swarm of up to 100 sub-agents, executing parallel workflows across up to 1,500 coordinated steps, without predefined roles or hand-crafted workflows. PARL uses a trainable orchestrator agent to decompose tasks into parallelizable subtasks, each executed by dynamically instantiated, frozen subagents. Running these subtasks concurrently significantly reduces end-to-end latency compared to sequential agent execution.

still cannot reach nano nanana pr

does it support ralph? harness? toolcalling? and all other modern buzzwords??? kind of excited to run it locally.

I was vibe testing K2.5 for this video earlier this week, and I’d never seen AI produce interactions like this before. So I used some of Kimi’s own creations as transitions in it.

Kimi K2.5 dropping with Aesthetic Coding and Agent Swarm is exciting.

How to understand the concept of the "Agent Swarm"?

Kimi k2.5 has arrived

Hi team. I can't seem to get the built-in web_search tool to work with the Kimi K2.5 model. Can someone take a look? It works with Kimi K2 thinking & kimi-k2 preview models.

Agent swarm looks promising

ok but can it center a div?

The AI agent era!

Kimchi?

niu

Unleash the Swarm. 🐝

Did you make it using Skills?

Huge thanks to the whole crew behind Kimi K2.5 crushing the drop!

Well done. I am interested in testing its visual understanding, integrating it with I will share a demo soon.

Congrats. As the Chinese Spring Festival approaches, more and more Chinese AI models are being released.

@grok can I use it for free ?

@bdsqlsz @grok 总结一下视频的核心内容

aesthetic coding unlocking agent swarms feels like a force multiplier

Amazing K2.5 leads open-source but holds its own closed-source: Edges GPT-5 on agentic tasks (HLE: 50.2% vs 41.7%) and SWE-bench coding (76.8% SOTA). Vs Claude 4.5: K2.5 beats on tools-enabled reasoning (44.9% vs 32%), but Claude wins MMLU-Pro (87.1% vs 84.6%). Vs Gemini 2.5 Pro: K2.5 crushes vision/coding (MMMU Pro 78.5%), but Gemini's 2M context dwarfs K2.5's 256K.

Still Open source?

Agents overtaking! Congrats on release @Kimi_Moonshot 🔥

Starting with Kimi, the 2026 AI Theater will officially begin performances.

Kimi has always been great writing. I’m curious about how 2.5 performs

Creative text like minimax-her? 🥺

So, no linear attention mixed in, as rumored. self.language_model = DeepseekV3ForCausalLM(config.text_config)

The agent swarms must be put into containers for a safer future world:

How did I only just hear about this? like, right now?

Love how the editing goes 0 to 100 instantly. The editor really said 'buckle up'.

There's truly no stopping the progression towards the ASI development. 👍


