MoonshotMoonshotΒ·text

Kimi K2.6

CodeVisionReasoningWeb searchFunction calling
Advertised as kimi-k2-6kimi-k2.6moonshotai/kimi-k2.6
Quick reference
Kimi K2.6 β€” TLDR
  • 🧠 Trillion-parameter MoE, only 32B active per token^2
  • πŸ†• Native multimodal agentic model from Moonshot AI^1,2
  • πŸ“ 256K context window, native INT4 quantization^2
  • πŸ‘οΈ Vision integrated via 400M MoonViT encoder^2
  • πŸ”§ Built for agent-swarm orchestration and tool use^2
  • πŸ’¬ Thinking and non-thinking modes; preserves reasoning^1
  • πŸ“š Open weights under Modified MIT License^2
  • 🎯 Targets long-horizon coding and autonomous execution^2
πŸ’° Best price on Antseed
$0.020 / $0.094βˆ’97%
per 1M Β· cheapest in / out
🐜 Providers
23
advertising on Antseed
πŸ”¬ Model checks
Not audited yet
no audit of this model so far
πŸ“ Context
256K tokens
Provider

Moonshot is an AI research lab known for developing the Kimi family of large language models. The organization has gained recognition for building capable reasoning-oriented models, with the Kimi line representing its flagship series of text generation systems.

Explore 8 more models by Moonshot β†’
About this model

Kimi K2.6 is Moonshot AI's open-weight, native multimodal agentic model, released April 20, 2026 and built on a sparse Mixture-of-Experts architecture.^2 It carries 1 trillion total parameters but activates only about 32 billion per token.^2 Vision is integrated architecturally through a 400M-parameter MoonViT encoder, letting the model accept text and images.^2 It exposes a 256K-token context window and ships with INT4 quantization.^2

Compared with its same-family predecessor Kimi K2.5, K2.6 retains the trillion-parameter MoE design and 256K context window while shifting its focus toward long-horizon coding, coding-driven design, and agent-swarm orchestration. Both belong to Moonshot's Kimi K series, and K2.6 is positioned as the newer iteration aimed at proactive, autonomous multi-step execution.

The model supports function calling, web search, and a mode that preserves reasoning across multi-turn interactions, alongside separate thinking and non-thinking responses.^1 Its open weights are distributed under a Modified MIT License,^2 and it is available through Hugging Face and inference providers.^1,2

Within the broader lineage, Moonshot later shipped the coding-focused Kimi K2.7 Code in a separate family. K2.6 remains aimed at developers who need open weights, a large context window, and large-scale agent orchestration.

View source on GitHub β†—View model card on HuggingFace β†—
Sources
build.nvidia.comkimi-k2.6 Model by MoonshotaiΒ· build.nvidia.comdocs.api.nvidia.commoonshotai / kimi-k2.6Β· docs.api.nvidia.comhuggingface.comoonshotai/Kimi-K2.6 Β· Hugging FaceΒ· huggingface.co

This About section is AI-generated from public sources via VeniceStats + Venice inference, with no human editing. It may contain inaccuracies.

Usage on AntseedΒ Β·Β Kimi K2.6
Tokens served
373.44M
input + output
Requests
20,183
settled calls
Buyers
90
distinct, on this model
Providers used
28
of 23 advertising
Settled
$325.88
gross USDC, this model

Daily usageΒ Β·Β Kimi K2.6Β Β·Β Organic providers

Top providersΒ Β·Β Kimi K2.6Β Β·Β Organic

Providers serving Kimi K2.6 (23)compare on the network explorer β†’
ProviderHealthRep↓SalesTokensRoutingInput $/MCached
$/M
Output $/MSave1st tokentok/s

"Best price" and the provider table are live Antseed catalog data (advertised $/1M tokens β€” or $ per generated image for unit-billed image models β€” not settled amounts). Reputation = buyer trust score (0-100, the Antseed SDK's own formula). "Routing" = the SDK's default buyer routing (what Antseed Desktop ships with): a trust β‰₯ 60 gate on the effective reputation, then cheapest-first among routable providers; live failover state (per-peer cooldowns) is buyer-side runtime and not included. Model knowledge (TLDR, provider, About) via the VeniceStats enrichment layer. Advertised catalog, not the model used in any specific purchase. "Usage on Antseed" counts only settlements whose buyers share the per-model split on-chain (metadata v2/v3, opt-in), so every usage figure is a lower bound.