Moonshot Kimi Blog
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Moonshot open-sources Kimi K2: a 1-trillion-parameter MoE built for agents
Moonshot AI released the Kimi K2 weights: 1 trillion parameters, 32 billion active, trained on 15.5 trillion tokens without a single failure. The model is built for tool use.
Moonshot releases open Kimi K2 Thinking: 300 tool calls in a row
Moonshot AI's open Kimi K2 Thinking scores 44.9% on HLE with tools, 60.2% on BrowseComp and 71.3% on SWE-Bench Verified, running natively in INT4.
Moonshot releases Kimi Vendor Verifier: 6 tests for checking K2.6 providers
With Kimi K2.6, Moonshot open-sourced KVV, a set of six checks for open-weight providers. A full run takes about 15 hours on two servers with 8 H20 GPUs each.
Moonshot releases Kimi K2.5: a swarm of 100 subagents and 1500 tool calls
Moonshot unveils the open multimodal Kimi K2.5: further trained on 15T tokens, with a swarm of up to 100 subagents and up to 4.5x faster execution on complex tasks.
Moonshot Releases WorldVQA: 3500 Questions About the World That Trip Up Models
The Kimi team published WorldVQA, a benchmark of 3500 image-question pairs. On rare objects, top multimodal models often fall short of 50% accuracy.
Kimi launches Agent Swarm: up to 100 subagents and 1,500 tool calls
Moonshot built a swarm mode into Kimi K2.5: the model hires up to 100 subagents on its own, makes more than 1500 tool calls and, the company says, works 4.5x faster.
Moonshot open-sources Kimi K2.6: 300 sub-agents and 13 hours of autonomous coding
Moonshot open-sourced Kimi K2.6: the model spent 13 hours tuning an exchange engine, sped up Zig inference from 15 to 193 tokens/s, and scales its swarm to 300 agents.
Kimi releases PerceptionBench: none of 16 MLLMs reaches 60%
Moonshot released a 3000-question benchmark where the answer is visible to the eye. None of 16 top multimodal models scored 60%, and hallucinations are the weakest spot.
Moonshot releases Kimi K3 with 2.8 trillion parameters and a 1M-token context window
Moonshot unveils Kimi K3: 2.8 trillion parameters, 16 of 896 experts active per token, a 1M-token context window and $15 per million output tokens.