Kimi K3: The 1M-Context Agent Flagship, Hands-On

Kimi K3 is Moonshot AI's flagship, live in the Kimi app, CLI, and desktop. It ships in two variants: K3 Max ("Chat & Agent, flagship all-rounder") and K3 Swarm Max ("Massive search, batch processing, and more in one go"). It is a 2.8-trillion-parameter Mixture-of-Experts model with native vision and a 1-million-token context. Moonshot's tech blog now carries official benchmarks and API pricing ($3/$15 per million tokens), so the specs are confirmed, not leaked. The one piece still pending is the open-weights drop, due by July 27, 2026.

Is Kimi K3 out yet?

Yes. kimi.com leads with "Kimi AI with K3," and it runs in the app, CLI, and desktop. The open weights are the only piece not yet shipped, due by July 27, 2026.

Kimi K3 Max and K3 Swarm Max: the two variants

The in-app picker lists three options:

• K3 Max — "Chat & Agent, flagship all-rounder." The default, for most coding, reasoning, and assistant work.

• K3 Swarm Max — "Massive search, batch processing, and more in one go." The heavy variant for complex multi-agent workflows.

• K2.6 Fast — "fast chat, quick replies." The prior flagship, now the low-latency option.

K3 Max runs the everyday agentic loop; K3 Swarm Max fans one task across many parallel sub-agents.

Kimi K3 Agent Swarm

Agent Swarm is K3's headline feature, which is why the flagship ships a dedicated Swarm Max variant. It runs massive parallel search and batch processing so one session gets far more done, extending the swarm system from Kimi Work, which already coordinated hundreds of sub-agents.

Context is a paid membership tier

Moonshot gates context length by plan: 256K on Moderato, 1M from Allegretto up, set by a separate "Context Length" control in the picker. Your tier changes how much of a project K3 keeps in working memory, not whether you can use it.

It is a clean picture of where AI pricing is heading: the model is the CPU, the context window is the RAM, and your membership is the RAM configuration. It also explains K3's focus on coding, game, and 3D work, a 1M context exists to swallow a codebase, docs, logs, and task history at once, not to chat longer. Pick the 1M tier for large-repo coding, multi-file agent runs, or long sessions; 256K covers ordinary chat and short tasks.

The K3 specs

With the official tech blog out, most of what circulated as a leak is confirmed.

Only the weights remain, due July 27. Until then, no self-hosting.

Kimi K3 benchmarks

At max reasoning effort, K3 posts frontier-level scores on Moonshot's own suite, edging Claude Opus 4.8 on every published benchmark and trailing only the top proprietary models, Claude Fable 5 and GPT-5.6 Sol. These are vendor-reported.

Independently, Artificial Analysis ranks K3 #4 of 189 on its Intelligence Index (57, one point above Opus 4.8 at 56). K3 also posted BrowseComp 91.2 (best at release) and 56.0 on Humanity's Last Exam with tools. Full head-to-head: Kimi K3 vs Claude Opus 4.8.

Caveat: the table is Moonshot's own, and Artificial Analysis flags K3 as slow and verbose, so weigh the scores against throughput.

Hands-on

From early use of K3 Max in Kimi Code (July 2026):

• Max thinking only. There is no lower-effort setting yet, so responses are slow. Quality holds up: on a real coding task it felt Opus 4.8-tier, which the benchmarks back.

• Time-based quota, generous. A four-hour window dropped only ~15% running a large task. K3 is built to be left working, not pinged for one-liners.

• Long single runs. Asked to build a native Mac interface for a networking tool, it settled into a job over an hour long, exactly the long-horizon work it targets.

• Nicer interface. Kimi Code is cleaner than before, down to a rotating-moon loading animation during the long thinks.

Kimi K3 vs DeepSeek V4 Pro, GLM 5.2, and Claude

K3's 2.8T parameters are pitched against DeepSeek V4 Pro (1.6T) and Baidu's ERNIE 5.0 (2.4T). Size is a poor proxy for quality, but it signals Moonshot chasing the frontier over efficiency. For the cheaper open options, see DeepSeek V4 Pro versus Flash and GLM 5.2 versus DeepSeek V4 Pro. Against closed models the reference is Claude, which K3's coding builds are already benchmarked against.

How to access Kimi K3

• Kimi's platforms — the app at kimi.com, CLI, and desktop (1M context on higher tiers), plus the API on Moonshot's Open Platform at $3/$15 per million ($0.30 cached). See the full pricing breakdown.

• Open weights — earlier flagships shipped on Hugging Face; K3's are due there by July 27, 2026.

• Aggregators — multi-model gateways add new Chinese open models fast; platforms like AIReiter already carry DeepSeek V4 and GLM 5.2 via an Anthropic-compatible API.

FAQ

What is the difference between K3 Max and K3 Swarm Max?

K3 Max is the general-purpose flagship for chat and everyday agentic tasks. K3 Swarm Max is the heavy variant for massive parallel search, batch processing, and complex multi-agent workflows.

Do I need the 1M context tier?

Only for work that needs a lot resident at once: large-repo coding, multi-file agent runs, or long sessions. The 256K Moderato tier covers ordinary chat and short tasks; the 1M window is on Allegretto and up.

Are Kimi K3 open weights available?

Not yet. Moonshot says the full weights (Modified MIT) ship by July 27, 2026; they are not on its Hugging Face org yet.

What is Kimi K3 Agent Swarm?

Its headline capability: coordinating many sub-agents to search and process in parallel so one session does far more. It extends the swarm system from Kimi Work.