Kimi K3 Explained: Moonshot AI's 2.5T Flagship Model — Leaked Specs, Verified Tech, and What to Watch

On July 15, 2026, Moonshot AI posted a 30-second teaser video. No specs, no benchmarks, no model card. Within hours: 620,000 views, a leaked billing page showing a top-up promotion running through August 11, and Polymarket odds hitting 97% for a this-month release.

The model is called Kimi K3. The leaked claims: 2.5 trillion total parameters, a native million-token context window, and pricing that would undercut every frontier competitor. Early testers on X are comparing it to GPT-5.6 and Fable 5.

The question is not whether K3 is coming — it is whether the claims hold up.

Important disclaimer upfront: Kimi K3 has not been released. As of July 16, 2026, there is no model card, no benchmarks, no license. This article is a pre-release explainer based on leaks, one verifiable technology (Kimi-Linear), and pre-launch signals. We label every claim by confidence level. Do not make production decisions based on this article — but do prepare to evaluate K3 the moment it goes live.

---

The Leaked Specs, Sorted by Confidence

Here is everything the leaks claim, organized so you know what to trust:

Rule of thumb: If an article quotes specific K3 benchmark scores, close the tab. No one has tested this model. Any numbers being cited are fabricated or extrapolated from other models.

Moonshot AI's track record adds credibility to the scale claims. K2.6 (~1T MoE) and K2.7 Code were both open-weight models that gained real developer adoption. K2.7 Code was officially described as "open-source multimodal." Moonshot has shipped at this level before — K3 at 2.5T would be a large jump, but not from a team without history.

---

What Million-Token Context Actually Means in Practice

Most frontier models today max out at 128K–200K tokens of context. K3 reportedly supports 1M natively — roughly 5–10× more.

Why does this matter? Because many real-world tasks exceed 200K tokens once you include all the relevant context:

• Full-codebase reasoning. A medium-sized codebase — repo structure, source files, documentation, test suites, commit history — easily exceeds 200K tokens. A 1M-token model could hold the entire repo in a single call, enabling codebase-wide refactoring, bug hunting, and architecture reviews without chunking.
• Long-document analysis. Legal contracts, research paper collections, financial reports, technical documentation — these regularly exceed what current models can hold. A million-token window means fewer API calls, no summarization losses, and better cross-reference accuracy.
• Multi-step agent workflows. Agent loops that iterate over many rounds accumulate context quickly. A 1M window means the agent can retain its full history — every decision, every observation, every correction — without truncation or summarization.

Rule of thumb: If your typical prompt is under 10K tokens, K3's 1M window is irrelevant to your workflow. If you routinely bump against 128K limits and resort to chunking, summarization, or multi-call workarounds, long context is the single most impactful feature to watch.

But Does Long Context Actually Work?

This is where Kimi-Linear — the one verifiable technology in this whole story — matters.

Many models claim long context but degrade in practice. They "support" 1M tokens but start forgetting or hallucinating past 200K. Kimi-Linear is an architectural solution, not a marketing trick:

• Most attention layers use a linear approximation that compresses the context into a fixed-size state, cutting KV cache memory by ~75%
• A minority of layers (every 4th to 8th) retain full quadratic attention for precise long-range lookups
• Published result: ~6× faster decoding at 1M tokens, with retrieval accuracy only 2–4 percentage points below full attention

The code and paper are on GitHub (MoonshotAI/Kimi-Linear). You can read the paper's Table 2 right now to see the throughput and accuracy numbers at different sequence lengths. This is not "trust us" — it is "here is the code, run it yourself."

---

Fast Decoding: Why 6× Matters for Iteration

Context length gets the headlines. Decoding speed is what you feel in your workflow.

Standard transformer attention has a quadratic relationship with sequence length. Double the context, and you roughly quadruple the compute for attention. At 1 million tokens, this creates two problems: extreme latency per generation, and extreme cost per generation.

Kimi-Linear's 6× decoding speedup attacks both:

For interactive use, the difference between 3-minute and 30-second responses is the difference between "usable" and "unusable." For agent loops, it is the difference between a pipeline that runs overnight and one that finishes during a coffee break.

---

The Multimodal Question

Multimodal support — the ability to process images, audio, and potentially video alongside text — is the most important unresolved question about K3.

If K3 is multimodal, it can reason directly over visual and audio input. This unlocks use cases like document understanding with images, UI analysis, audio transcription with reasoning, and mixed-media agent workflows.

If K3 is text-only, it remains a powerful language and reasoning model but needs separate vision or audio models piped in for non-text tasks.

The leak evidence is genuinely conflicting. Some sources say multimodal, others do not mention it. K2.7 Code was multimodal, which means Moonshot has the capability — but they may have scoped K3 differently.

Rule of thumb: Assume text-only until proven otherwise. If multimodal arrives, treat it as a bonus. Do not build workflows that require K3 multimodal and then scramble if it launches text-only.

---

Cost Modeling: What the Leaked Pricing Implies

If the leaked pricing is directionally correct, K3 would be dramatically cheaper than current frontier models for high-volume use:

At leaked pricing, K3 would reduce API costs by 90–95% compared to Claude Opus for equivalent workloads. For teams running thousands of API calls per day — agent pipelines, batch processing, RAG systems — that cost difference compounds fast.

Caveat: These numbers assume the leaked pricing is accurate (low confidence) and that K3's output quality matches frontier models (zero evidence). A cheaper model with worse output quality just means more iterations and more human review, which can actually increase total cost. Do not commit budget until you test.

---

How K3 Compares to the Field

Without official benchmarks, we can only compare architecture and known facts:

If K3 ships as open-weight with verified 1M context and performance between GPT-5.6 and Fable 5 (as early testers claim), it would be the strongest open-weight model available. But early tester claims are not benchmarks — wait for Artificial Analysis, LMSYS Arena, or SWE-bench Verified results.

---

What to Do Before Launch

You do not need to wait for the release to start preparing:

Before launch (now):

• Read Kimi-Linear's paper Table 2 on GitHub — form your own opinion on whether 1M context is architecturally viable (~20 minutes)
• Identify the 2–3 workflows where context overflow or API costs currently cause you the most pain
• Write your hardest test prompt — the one that requires the most context and the most iterations — so you can run it immediately at launch

Launch day:

• Run your test prompt on K3. Compare the output to the same prompt on your current model (Claude, GPT, DeepSeek)
• Specifically check: does quality hold at long context? How fast is each iteration? Does it handle complex multi-step reasoning?
• Do not read the official blog post benchmarks — test with your own tasks

Two to four weeks after launch:

• Wait for independent evaluations (SWE-bench, LMSYS Arena, Artificial Analysis)
• If open weights arrive, evaluate whether fine-tuning on domain-specific data is viable
• Re-run the cost model above with real pricing and your actual usage pattern

---

FAQ

Is Kimi K3 released?

No. As of July 16, 2026, Kimi K3 has not been officially released. Moonshot AI posted a teaser video on July 15 (620,000+ views), a leaked billing page suggests an imminent launch, but there is no public API access, model card, or benchmark report.

What are the leaked Kimi K3 specs?

Approximately 2.5T total MoE parameters, 40–60B active parameters per token (estimated), 1M native context via Kimi-Linear architecture. No official specs have been confirmed.

How much will Kimi K3 cost?

Leaked estimates: ~$0.80–1.20 per million input tokens, ~$3.50–4.50 per million output tokens. Source is a single leaked billing page — treat as directional, not confirmed.

Will Kimi K3 be open source?

Unknown. Moonshot's previous models (K2.6, K2.7 Code) were open-weight, which is a positive signal. However, rumors suggest K3 may launch API-first with open weights delayed to Q4 2026.

How does Kimi K3 compare to GPT-5.6 or Fable 5?

No performance comparison is possible until K3 has benchmarks. Architecturally, K3's advantages are its 1M native context (versus 128K–200K for GPT and Claude) and potentially much lower pricing. Early testers claim performance between GPT-5.6 and Fable 5, but no independent evaluation has confirmed this.

Is Kimi K3 multimodal?

Unconfirmed. Leak evidence is conflicting — some sources say it supports text, image, and audio input; others do not mention multimodal at all. K2.7 Code was multimodal, so Moonshot has the capability, but K3's modality support is unknown until official specs arrive.

---

Bottom Line

Every few months, a new model promises to change everything. Most do not.

What makes K3 worth watching is not the parameter count or the hype. It is two specific, testable claims: million-token context that actually works (backed by published, open-source architecture), and a 6× decoding speedup that makes long-context use practical rather than theoretical.

None of this is proven at the model level yet. K3 is vaporware until the API opens and real users test it on real tasks.

But the underlying technology (Kimi-Linear) is published and testable today. The pricing signals, if directional, would make high-volume API use viable at scales that are currently too expensive. And the launch window looks like days, not months.

So here is the smallest useful action you can take right now: identify the workflow where you currently hit context limits or cost walls, write down your hardest test prompt, and have it ready for launch day. Everything else is noise until you can run real prompts against a real model.