Kimi K3: 2.8 Trillion Parameters
and What It Actually Changes for Your AI Stack

On July 17, 2026, Moonshot AI announced Kimi K3 — a 2.8 trillion parameter model. The previous record among open-weight models was DeepSeek V4 Pro at 1.6T (April 2026). K3 is 75% larger.

Within 48 hours of the announcement, demand pushed close to the limits of Moonshot's current GPU capacity. On July 19, the company paused new subscriptions and announced that full model weights will be released by July 27 under a Modified MIT license.

This is not just another large model. It is a moment when open-weight AI stops being a compromise between power and price — and becomes a real infrastructure alternative. Here is what actually changed and what to do about it in the next seven days.

1. What happened

Moonshot AI — a Chinese AI startup founded in 2023 — released Kimi K3. The facts that matter:

Architecture. 2.8 trillion total parameters, Mixture-of-Experts design: 16 out of 896 experts activate per inference pass. Moonshot's official blog confirms the 16/896 structure but has not published an official active parameter count. Community estimates put active params at approximately 50B (calculated as 2800/896 × 16). Use that figure as a community estimate, not an official specification.

Context window. 1,048,576 tokens — exactly 1 million. For comparison: Claude 3.5 Sonnet supports 200K, GPT-4o supports 128K. K3 can process an entire mid-size codebase, a year of customer correspondence, or a full legal document archive in a single inference call.

Multimodality. Native support for text and images is confirmed by the official blog and multiple independent sources. Video is mentioned in one source only and should not be treated as a confirmed specification.

API pricing. $3 per 1M input tokens, $15 per 1M output tokens. GPT-5.6 Sol runs at $5/$30. K3 is approximately 40% cheaper on input and 50% cheaper on output compared to OpenAI's flagship tier. Important caveat: GPT-5.6 Terra at $2.50/$15 is cheaper than K3 on input. The comparison holds only against Sol-tier, not the full GPT-5.6 lineup.

Availability. Moonshot paused subscriptions on July 19 due to capacity pressure and announced two new plans: Kimi Membership (web, mobile, workplace) and Kimi Code Membership (coding workflows). The API is available. Full weights release July 27 under Modified MIT.

2. Why this is a paradigm shift

Before K3, the mental model was clear: closed models (GPT-5.6, Claude 4) offered maximum capability at premium cost; open models (Llama, DeepSeek) offered a tradeoff — cheaper but weaker on complex reasoning tasks.

K3 disrupts this on three axes simultaneously.

3. The new architecture in plain English

Mixture-of-Experts is a way to scale a model without proportional growth in inference cost.

Think of a library with 896 specialists. Each incoming request goes to 16 of them — whichever the router considers most relevant. The other 880 experts consume no compute during that inference pass.

The result: total model size of 2.8T parameters gives it broad knowledge and deep specialization across domains. But each individual inference call uses approximately 50B active parameters (community estimate based on 16/896 structure). This means K3 does not require GPU memory for 2.8T parameters to process a single token — only memory for the active experts plus the routing mechanism.

4. My Content Factory case: real numbers

I run Content Factory — a pipeline that takes a topic, processes it through multiple models in parallel for different tasks, and produces content for 15+ distribution channels. Current stack: Claude API for structure and fact-checking, Gemini for tone and long-form generation, Groq for speed-sensitive micro-tasks.

The primary bottleneck in this pipeline is not quality — it is the cost of long-context inference. When I run fact verification against a full brief (5-10K tokens) plus source material (another 15-20K tokens), each pass costs $0.10-0.20 using premium models. At 50 passes per day, that is $5-10 per day on fact-checking alone.

Kimi K3 at $3/1M input is approximately 1.5-2× cheaper than Claude 3.5 Sonnet on long-context tasks for comparable workloads. I am not publishing specific savings numbers here because quality on my specific tasks needs to be tested — that test runs this week.

More important for my architecture: 1 million tokens of context means I can pass the entire publication archive from the last three months in a single request and ask the model to identify patterns in what performs well. Currently that requires either chunking and aggregating answers, or building a separate RAG system. With K3 — potentially one call. If the July 27 weights come out with clean Modified MIT licensing, I will evaluate self-hosted deployment for tasks where client data cannot go through external APIs.

5. The cost math that wakes up CFOs

Direct cost comparison for a team of 5 developers, 200 API calls per day, average 2K input / 1K output tokens per call:

Difference: $114/month at the same request volume. At 10× enterprise load (2,000 calls per day): $1,140/month in API cost savings.

One important context for the CFO framing: K3 is cheaper than GPT-5.6 Sol, but GPT-5.6 Terra ($2.50/$15) is in a comparable price range and cheaper on input. The economic case for K3 over Terra is the context window advantage, not pure cost. If your tasks fit within 128K-200K tokens, Terra or Luna may still be the rational choice. Build a mixed stack rather than all-in on any single model.

6. What dies, what lives

7. What to build this week

8. B2C / B2B split

Frequently Asked Questions