Logo
Moonshot · Kimi K3 · Jul 2026

Kimi K3 — long-horizon coding flagship with million-token context

  • Million-token context
  • Native vision input
  • Deep reasoning built in
  • Long-horizon Agent focus
Kimi K3

What is Kimi K3?

Kimi K3 is Moonshot’s open frontier flagship, released in July 2026, for long-horizon coding, analysis and writing, and Agent workflows. Versus Kimi K2.6, context rises to a million tokens, with native vision and built-in, tunable deep reasoning. For everyday high turnaround, compare lighter models in the family; for top closed-source peers, look at Claude Opus 5 or GPT-5.6 Sol. Pick Kimi K3 in iMini Agent to use it.

Vendor
Moonshot
Released
Jul 2026
Context
1M tokens
Max output
Per platform limit
Deep reasoning
Built-in · tunable
Best as
CN/EN long-context flagship workhorse

What’s new vs Kimi K2.6

Context, vision, and deep reasoning—the changes to check before you make Kimi your flagship pick.

Context up to a million tokens

Versus K2.6’s ~256K, long specs and large-repo material can stay in one thread.

Native vision input

Images enter the same reasoning loop—handy for coding or analysis with screenshots, charts, and UI.

Deep reasoning built in

Raise the reasoning tier on hard jobs; keep a lower tier for everyday speed.

Stronger long-horizon Agent focus

Built for coding, research, and multi-step tool work—keeping the goal until you have a shippable result.

Official benchmarks

Charts from Moonshot’s Kimi K3 technical blog. All peers were evaluated at max / xhigh thinking strength.

Kimi K3 internal knowledge-work benches: Online Exp, DECK, Finance
Three internal knowledge-work scores. Online Exp Bench 75.5, DECK-Bench 73.5, Finance-Bench 62.6—ahead of GPT-5.5 and Claude Opus 4.8 on the same board. Knowledge synthesis and finance materials are among K3’s strengths.
Kimi K3 Agent / tool-use related benchmarks
Agent and tool-related benches. Multi-step tool calls, browsing, and automation—use these to judge whether K3 finishes real workflows, not just single-turn Q&A.
Kimi K3 vision and document-understanding benchmarks
Vision and document understanding. Native multimodal peers; keep image, table, and long-doc scores separate from pure-text marks.
Kimi K3 reasoning benchmark comparison
Reasoning-suite comparison. Shown with Claude Fable 5, GPT-5.6 Sol, and peers; K3 leads among open weights but still trails the strongest closed-source tiers.
Kimi K3 coding benches: DeepSWE, FrontierSWE, Terminal Bench, and more
Six coding scores. DeepSWE 67.5, FrontierSWE 81.2, Terminal Bench 2.1 88.3, Program Bench 77.8, SWE Marathon 42.0—several ahead of Opus 4.8, still below Fable 5 / GPT-5.6 Sol at the top.
Kimi K3 supplemental benchmark chart
Supplemental benches—read with the main tables. Method notes cover harness and peer sourcing; across vendors, trust the method first.

Three common workflows

Kimi K3 fits long materials and long-horizon jobs in CN/EN settings.

Long-horizon coding

Cross-file features, test fixes, and repo-scale changes. Raise reasoning when the job is hard.

Knowledge work with images

Read screenshots, organize charts, and write decision memos—cite what you can see.

Multi-step Agents

Tool use and research-style automation. Agree success checks and stop conditions up front.

Vs Kimi K2.6 and GLM 5.2

All three lean Chinese and open-weight paths. The gaps are context, vision, and flagship generation.

DimensionKimi K3Kimi K2.6GLM 5.2
Context1M tokens~256K tokens1M tokens
ReasoningDeep · tunableMulti-subagent orchestrationThinking High / Max
Coding / AgentsLong-horizon coding flagshipLong-horizon coding & multi-subagentsOpen coding / agent flagship
Long tool runsVision + long threadsMulti-subagent orchestrationLong-horizon delivery focus
Speed postureLower tier for speedHeavier long runsTunable thinking budget
Prefer whenNew Kimi flagship jobsWorkflows still on K2.6Tied to the Zhipu stack

Using Kimi K3 on iMini

Free quota, model identity, key specs, and how it differs from K2.6 and GLM.

More models

More flagship models by vendor

GPT-5.6 SolOpenAI’s depth flagship—when the team is on the GPT stack and needs maximum depth.Claude Opus 5Anthropic’s Opus flagship—for long-horizon coding and multi-step Agents on the Claude stack.Gemini 3.1 ProGoogle’s Gemini Pro tier—for multimodal, long-context work on the Google stack.Grok 4.5xAI’s Grok flagship—when you need Grok reasoning or the xAI product stack.DeepSeek V4 ProDeepSeek’s coding and reasoning flagship—a strong open-weight pick for code and hard reasoning.GLM 5.2Zhipu’s GLM flagship—for bilingual Agents and Zhipu-stack delivery.Qwen3.7 MaxQwen Max—when you need peak Qwen text reasoning.MiniMax M3MiniMax’s flagship—when delivery is tied to the MiniMax stack.Llama 4 MaverickMeta’s Llama 4 flagship option—when you need the Llama weight stack.Seed 2.0 LiteByteDance Seed’s lighter tier—for high-turnaround, shorter sessions.
All models

Use Kimi K3 free on iMini Agent

Million-token context and native vision—for long-horizon coding and knowledge work.

Open Kimi K3

No Moonshot API key required.