Kimi K3: What Moonshot’s 2.8T Flagship Means for Builders

Moonshot AI’s Kimi K3 is the lab’s most capable flagship to date: a 2.8-trillion-parameter open 3T-class model with a 1-million-token context window and native vision. If you build agents, ship frontend/product code, or do long research sessions, here is what matters in practice — what it is good at, what it costs, and how to adopt it without getting surprised by the launch constraints.

Sources for the facts below: Moonshot’s Kimi K3 tech blog and the official Kimi K3 quickstart.

What Kimi K3 is

Kimi K3 (moonshotai/kimi-k3 on OpenAI-compatible gateways; kimi-k3 on Moonshot’s own API) is a sparse Mixture-of-Experts model that Moonshot describes as:

• 2.8T total parameters, effectively activating 16 of 896 experts
• Built on Kimi Delta Attention (KDA) and Attention Residuals
• Native visual understanding (text + image input; text output)
• 1,048,576-token context with flat, non-tiered context pricing
• Aimed at long-horizon coding, knowledge work, and deep reasoning

Moonshot is explicit that overall performance still trails the strongest proprietary systems it compares against (Claude Fable 5 and GPT-5.6 Sol), while claiming frontier-level results across its evaluation suite and strong gains versus earlier open models. Treat vendor benchmarks as directional; the useful takeaway is the job family K3 is optimized for: multi-hour engineering sessions, large repos, tool use, and research workflows that mix text with screenshots.

Full open weights are scheduled by July 27, 2026, with a fuller technical report promised alongside that release. Until then, API and product surfaces are the practical path.

Specs that affect real work

Property	Kimi K3
API model id (gateway)	moonshotai/kimi-k3
Official Moonshot id	kimi-k3
Total / active experts	2.8T MoE; 16 of 896 active
Context	1,048,576 tokens
Modalities	Text + image in → text out (official API also documents video via file upload)
Thinking	Always on; reasoning_effort: low / high / max (default max at launch)
Max completion tokens	Default 131,072; up to 1,048,576
Sampling	Fixed at launch (temperature=1.0, top_p=0.95, etc. — omit overrides)

Two integration details from the official docs are easy to miss:

1. Always return the full assistant message in multi-turn and tool loops — including thinking history. K3 was trained with preserved thinking history; dropping it (or mid-session switching from another model) can make quality unstable.
2. Vision on the official API does not accept arbitrary public image URLs; use base64 or Moonshot’s ms:// file references, and send content as an array of parts.

Pricing: where the bill actually lands

Official Kimi API list pricing for K3 (flat across the 1M window):

Token type	Price
Cache-hit input	$0.30 / MTok
Cache-miss input	$3.00 / MTok
Output	$15.00 / MTok

Moonshot reports >90% cache hit rates on coding workloads on its infrastructure when long prefixes stay stable. That is the main cost lever: keep system prompts, repo digests, and tool schemas unchanged across turns so later requests can hit cache. Output is the expensive side — long agent traces with verbose reasoning at max effort add up quickly.

On this platform, moonshotai/kimi-k3 is already listed; check the live retail quote on Pricing before you budget, since gateway retail can differ from list.

When K3 is the right call

Reach for K3 when you need:

• Long-running coding agents that navigate large repositories and terminal tools
• Frontend / product work that benefits from “vision in the loop” (screenshots → revise UI)
• Research and knowledge-work agents that produce structured reports, charts, or long multi-step analyses
• A frontier-capable open-weight path once the July 27 weights land (self-host or partner inference)

Prefer a cheaper Kimi or another model when you need:

• Short chat, high QPS, or casual autocomplete — K2.x-class models are usually enough
• Strict “do nothing without permission” agents — Moonshot notes K3 can be over-proactive on ambiguous tasks; constrain it in the system prompt / AGENTS.md
• Zero-surprise sampling controls — K3’s fixed sampling at launch may not fit every eval harness

Practical adoption checklist

• Start with reasoning_effort="high" (or low) for cost-sensitive paths; reserve max for hard jobs after you measure quality vs. spend.
• Design for prompt caching: one stable long prefix, many short questions.
• Use an OpenAI-compatible client and set model to moonshotai/kimi-k3 if you already route through a unified gateway:

curl https://openfluxhub.com/v1/chat/completions \
  -H "Authorization: Bearer $YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshotai/kimi-k3",
    "messages": [{"role": "user", "content": "Summarize the risks in this migration plan."}]
  }'

• Keep tool-call loops faithful: append the complete assistant message, then one tool result per tool_call_id.
• If you self-host later, Moonshot recommends supernode-style deployments with 64+ accelerators — plan infra separately from API trials.

Bottom line

Kimi K3 is Moonshot’s bid for open frontier intelligence: 2.8T-scale MoE, 1M context, native vision, and API access now, with weights promised by July 27, 2026. It is not positioned as the single best closed model in the world — Moonshot says as much — but it is a serious option for builders who need long-horizon coding and knowledge-work agents at a transparent $3 / $15 token price with aggressive cache discounts.

moonshotai/kimi-k3 is live on this platform. Open the model page, create a key, and run it against one real repo or research brief before you change production defaults.