What this is
Moonshot AI's Kimi line went from a long-context chatbot to the best-in-class open-weight coding model in 12 months. Kimi K2.6 (April 2026) is the first open-weight model to beat GPT-5.4 (xhigh) on SWE-Bench Pro. This page is a 2026-05-15 release-by-release reference.
Kimi Release Timeline
| Date | Release | Notable |
|---|---|---|
| Pre-2025 | Original Kimi chatbot | Long-context (200K+) consumer assistant |
| Jul 2025 | Kimi K2 (open-source debut) | 1T-parameter MoE, 32B active |
| Jan 2026 | Kimi K2.5 | Multimodal (vision + text), 256K context |
| Apr 13, 2026 | Kimi K2.6 Code Preview | Beta release to limited testers |
| Apr 20, 2026 | Kimi K2.6 GA | 1T MoE, 32B active, 262K context, 300-agent swarms |
| Jun-Jul 2026 (expected) | Kimi K3 | Reportedly 3-4T parameters |
Kimi K2.6 Specifications
| Spec | Value |
|---|---|
| Total parameters | 1,000B (1T) |
| Active parameters per token | 32B |
| Architecture | MoE: 384 experts (8 selected + 1 shared) |
| Layers | 61 |
| Attention heads | 64 (Multi-head Latent Attention) |
| Context window | 262,144 tokens |
| Agent Swarm | Up to 300 sub-agents, 4,000 coordinated steps |
| SWE-Bench Pro | 58.6 (beats GPT-5.4 xhigh at 57.7) |
| License | Open weights |
Six Things the Lineage Tells You
- Moonshot is the fastest-climbing open-weight lab. From product-only in late 2024 to leading open-weight coding in mid-2026.
- Kimi K2.6 is the first open-weight model to beat GPT-5.4 on SWE-Bench Pro. 58.6 vs 57.7 — narrow but historic.
- Agent Swarm at 300 sub-agents, 4,000 steps is unmatched at this scale. Most open-weight agentic frameworks cap at 10s of sub-agents.
- Context window grew from 200K to 262K. Smaller jumps than competitors (Llama 4 Scout: 10M, Qwen3: 256K) but consistent.
- Multi-head Latent Attention (MLA) is the architectural signature. Same family as DeepSeek's efficient attention.
- K3 expected June-July 2026 targets 3-4T parameters. Direct scale match to the frontier closed labs.
What This Means for AI Visibility
Kimi K2.6 is increasingly deployed inside agentic coding workflows because of its SWE-Bench leadership. Brands selling developer tools, libraries, or APIs should test how their products surface inside Kimi-powered coding agents. The default model behind many emerging Chinese coding assistants and on-prem enterprise deployments is now Kimi K2.6 rather than DeepSeek V4.
Methodology
Release dates and specifications from LLM Stats Kimi K2.6 page, Kimi K2.6 release blog, Shanghai NYU RITS analysis, Cloudflare Workers AI changelog, and Miraflow's K2.6 analysis.
How Presenc AI Helps
Brand-visibility tracking on Kimi-powered surfaces (Kimi chatbot, Kimi K2.6 in agentic coding tools, on-prem K2.6 deployments) is increasingly important as the model absorbs share from DeepSeek and proprietary alternatives. Presenc AI runs the same prompt suites against Kimi K2.6 as against ChatGPT, Claude, and Gemini.