Pular para o conteúdo principal
Close
AI

KIMI 3: 2.8 Trillion Parameters in MoE That Shakes the LLM Podium

Gabriel Ferraresi· CEO | Tech86July 18, 20264 min
aillmkimi3moemoonshot-aiopen-weightslgpdsovereigntyinferenceagents

The heavyweight LLM table changed in July 2026. Moonshot AI released KIMI 3 on July 16, and the podium is no longer stable. We have been tracking this shift since the emergence of large-scale open-weight models — and the signal now is clear: $2 billion well placed opens space on the podium.

What KIMI 3 is: 2.8T in MoE, 16 active experts out of 896

According to Moonshot AI, KIMI 3 has 2.8 trillion parameters in a MoE (Mixture of Experts) architecture, with 16 active experts out of a total of 896 and a 1M token context window. Moonshot AI is a Chinese company founded in 2023 by Zhilin Yang, backed by Alibaba, that raised $2 billion.

MoE means only 16 experts run per token. Inference cost drops proportionally: instead of activating 2.8 trillion parameters, it activates only the fraction corresponding to the 16 selected experts. Comparable intelligence, a fraction of the compute. It is the architecture that lets it compete with top dense models without the inference cost of a 2.8T dense model.

The scoreboard on the benchmarks that matter

According to Moonshot AI's table, KIMI 3 scores 93.5% on GPQA Diamond — the best open-weight result ever published on that benchmark. It surpasses Claude Fable 5 (92.6%) and Claude Opus 4.8 (91.0%), sitting above both Anthropic models. GPT-5.6 Sol sits at 94.1% and still leads that specific benchmark — K3 still loses to OpenAI's top on GPQA Diamond.

According to Arena.ai, KIMI 3 ranks first on Code WebDev. Absolute leadership in code that runs in the browser. That is where the model shines most.

According to Moonshot AI, on HLE without tools KIMI 3 scores 43.5%. With tools, it rises to 56.0%. A 12.5 point difference when the model is given tools. A real agent.

According to the Artificial Analysis Intelligence Index, KIMI 3 is in fourth place among 189 models with 57.1 points. It still loses overall to Claude Fable 5 and GPT-5.6 Sol. But in coding and general agents, it edged past Claude Opus 4.8 and GPT-5.5.

Price and cost-performance: where DeepSeek V3 still wins

According to Moonshot AI, KIMI 3's price is $3.00/MTok input on cache miss, $0.30/MTok on cache hit, and $15.00/MTok output. The international API has been available at api.moonshot.ai since launch — anyone who wants to test it now can.

DeepSeek V3 is still cheaper. KIMI 3 does not play on minimum price — it plays on cost-performance. Top intelligence in coding and agents for a fraction of the compute of an equivalent dense model. Context caching is where the real saving shows up: $0.30/MTok on cache hit versus $3.00 on cache miss. Not using cache burns money.

What changes for Brazil: LGPD and data sovereignty

According to Moonshot AI, open weights land on July 27, 2026. That changes the game for Brazil. Self-host on-prem. Data sovereignty.

LGPD, Brazil's data protection law, requires security measures for personal data. With open weights, the model runs inside the client's datacenter. Personal data does not need to cross the border. No dollar-denominated API needs to enter the personal data flow. It is compliance by architecture, not by contract. It is data sovereignty in practice — and it is what sets KIMI 3 apart from closed models that only run on international endpoints.

What Tech86 implements with KIMI 3

We implement five workstreams with KIMI 3:

  1. On-prem inference via NVIDIA NIM with open weights. Sensitive data never leaves the client's datacenter. LGPD-compliant by architecture.
  2. Self-host open weights on an H100 cluster. Cost per token drops to hardware depreciation. No vendor spread, no lock-in.
  3. Prompt engineering for code agents. KIMI 3 leads Arena Code WebDev, according to Arena.ai. That is where it shines most.
  4. Context cache exploiting 1M tokens. $0.30/MTok on cache hit versus $3.00 on cache miss. Not using cache burns money.
  5. Routing between KIMI 3 and DeepSeek V3 by workload. Cheap for simple tasks, heavy for reasoning. Cost-performance optimization per token.

Conclusion: the LLM podium is no longer stable

Moonshot just proved that $2 billion well placed opens space on the podium. The LLM podium is no longer stable. KIMI 3 is not the best model overall — it still loses to Claude Fable 5 and GPT-5.6 Sol on the Artificial Analysis Intelligence Index. But in coding and agents, it sits at the top. And with open weights on July 27, 2026, it is the first podium-tier model Brazil can host locally without giving up data sovereignty. At Tech86, we help companies do exactly that — on-prem deploy, workload routing, and LGPD compliance by architecture.

Need expert guidance?

Schedule a consultation with our specialists.

AI Engineering with Data Sovereignty

Frequently Asked Questions

KIMI 3 is the model released by Moonshot AI on July 16, 2026. According to Moonshot AI, it has 2.8 trillion parameters in a MoE architecture, 16 active experts out of 896, and a 1M token context window. Moonshot AI is a Chinese company founded in 2023 by Zhilin Yang, backed by Alibaba, that raised $2 billion. The differentiator is cost-performance: comparable intelligence to top dense models, a fraction of the compute.

MoE stands for Mixture of Experts. According to Moonshot AI, KIMI 3 has 896 experts, but only 16 run per token. Inference cost drops proportionally: instead of activating 2.8 trillion parameters, it activates only the fraction corresponding to the 16 selected experts. Comparable intelligence, a fraction of the compute. That is how the model delivers 93.5% on GPQA Diamond at a price of $3.00/MTok on cache miss.

According to Moonshot AI's table, KIMI 3 scores 93.5% on GPQA Diamond — above Claude Fable 5 (92.6%) and Claude Opus 4.8 (91.0%). GPT-5.6 Sol sits at 94.1% and still leads that specific benchmark. According to Arena.ai, KIMI 3 ranks first on Code WebDev. According to the Artificial Analysis Intelligence Index, KIMI 3 is fourth among 189 models with 57.1 points — it still loses overall to Claude Fable 5 and GPT-5.6 Sol, but in coding and general agents it surpassed Claude Opus 4.8 and GPT-5.5.

According to Moonshot AI, open weights land on July 27, 2026. That enables self-host on-prem: the model runs inside the client's datacenter and personal data does not need to cross the border. No dollar-denominated API needs to enter the personal data flow. LGPD, Brazil's data protection law, requires security measures for personal data — hosting the model locally is compliance by architecture, not by contract. It is data sovereignty in practice.

According to Moonshot AI, the price is $3.00/MTok input on cache miss, $0.30/MTok on cache hit, and $15.00/MTok output. DeepSeek V3 is still cheaper. KIMI 3 plays in cost-performance: it is not the cheapest, but it delivers top intelligence in coding and agents for a fraction of the compute of an equivalent dense model. Routing between the two by workload is what optimizes total cost.

Blog — Get in Touch

Have a question about our articles or services? Our team is ready to help.

Schedule a Meeting

Book a time slot.

Schedule Now

Email

Send us a message.

[email protected]

WhatsApp

Quick conversation.

Address

Avenida Paulista, 1636 - São Paulo - SP - 01310-200

Tech86 Specialist

Online now

Hello! How can we help scale your business today?

Tech86 Engineering

We Value Your Privacy

We use cookies and similar technologies to optimize your experience, analyze site traffic, and personalize content. By clicking "Accept All", you agree to the use of all cookies. Read our Privacy Policy.