Kimi K3: Everything You Need to Know About Moonshot AI's 2.8-Trillion Parameter Open Frontier Model

Architecture

Kimi K3 is a 2.8-trillion-parameter sparse Mixture-of-Experts (MoE) model. Of its 896 total experts, only 16 are active per token, a design that aims to deliver the capability of a massive model with the computational cost of a much smaller one . Its key architectural innovations include:

• Stable LatentMoE: Moonshot's custom MoE formulation, which the company says improves training stability and efficiency .
• Kimi Delta Attention (KDA) and Attention Residuals (AttnRes): New attention mechanisms that Moonshot claims deliver up to 6.3x faster training and inference compared to standard attention mechanisms .
• 1-million-token context window with native vision capabilities, enabling the model to process long documents and images in a single prompt .
• At launch, K3 runs exclusively in a default max-effort reasoning mode; no faster or lighter inference option is yet available .

These architectural choices position K3 as a direct challenger to the most expensive frontier models, aiming to match their intelligence while offering substantially better value.

Benchmark Performance

Kimi K3's benchmark scores place it roughly between Tier-1 closed models and the open-weight pack. Key independent scores tell the story :

• Artificial Analysis Intelligence Index: 57.1 – ranks #4 overall, behind Claude Fable 5 (~60) and GPT-5.6 Sol (~59), and narrowly ahead of Claude Opus 4.8 (~56) .
• LMArena Frontend Code Arena: #1 – K3 took the top spot in blind tests for front-end coding, beating all other models including Claude Fable 5 and GPT-5.6 Sol .
• LMArena Chatbot Arena (StyleCtrl): #9 overall with an Elo of ~1380 .
• GPQA Diamond: scored 93.5%, the strongest open-weight result to date at launch, edging out Claude Fable 5's 92.6% on this PhD-level science reasoning benchmark .
• SWE-bench Verified: vendor-reported score of 80%+, indicating strong agentic coding performance .
• Terminal-Bench 2.1: scored 88.3%, beating Claude Fable 5 on this agentic terminal benchmark .
• GDPval v2: scored 1668 Elo on this knowledge-work leaderboard, surpassing Claude Opus 4.8 (1600) but trailing Claude Fable 5 (1760) and GPT-5.6 Sol (1748) .

In Moonshot's own 14-chart launch comparison, K3 beats Claude Fable 5 on 6 benchmarks: Terminal-Bench 2.1, Program Bench, SWE Marathon, Automation Bench, SpreadsheetBench 2, and BrowseComp. Fable 5 wins the remaining 8, with FrontierSWE showing the widest gap (86.6 versus K3's 81.2) .

Independent testers note that K3 is slower and more verbose than comparable models, and that only "max effort" inference mode is available so far . All benchmark comparisons currently rely on vendor-reported claims and third-party aggregations from Artificial Analysis and LMArena; independent peer-reviewed verification of the full suite is still pending .

Pricing

Kimi K3 API pricing as of July 16, 2026 :

• Input (cache miss): $3.00 per million tokens
• Cached input: $0.375 per million tokens (reported as $0.30 in several sources )
• Output: $15.00 per million tokens

This pricing is roughly 3–10x cheaper than Claude Fable 5 and GPT-5.6 Sol on a per-token basis . However, it is also approximately three to four times more expensive than Moonshot's previous flagship model K2, marking an end to the era of ultra-cheap Chinese API access . The price is flat across the entire 1-million-token window, not stepped by context tier .

Open-Weight Release Timeline

• API launch: July 16, 2026 at platform.kimi.ai .
• Full weights on Hugging Face: Promised by July 27, 2026, under a Modified MIT license that is commercial-use friendly but includes restrictions .
• As of the available reporting dates (through July 18, 2026), the weights are not yet public – only the API and a researcher-access program are live .
• Once released, K3 will be the largest open-weight model ever published, exceeding DeepSeek's V4 Pro (1.6 trillion parameters) and Zhipu AI's GLM 5 series (744 billion) by a wide margin .

The two-step rollout – paid API first, open weights eleven days later – is the defining shape of this announcement . Moonshot calls K3 "open frontier intelligence," but it is best described as open-weight in commitment rather than in practice until the checkpoint is downloadable .

Strategic & Funding Context

Moonshot AI has been on an extraordinary fundraising tear, driven by the need to keep pace in China's hyper-competitive AI race :

• November 2025: Raised funding at roughly a $4 billion valuation .
• February 2026: Closed a Series Unknown of over $700 million .
• May 2026: Raised ~$2 billion at a $20 billion valuation, led by Meituan's venture arm Long-Z Investments .
• June 2026: Opened talks for another $1–2 billion round targeting a $30 billion valuation – its third raise in six months .

Total disclosed funding is approximately $3.77 billion across 4 rounds . Backers include Alibaba, Tencent, Meituan, 5Y Capital, IDG Capital, and Zhen Fund . By April 2026, Moonshot's annual recurring revenue topped $200 million, driven by subscriptions to its Kimi chatbot and AI model services .

The strategic picture: Moonshot is burning capital aggressively to stay in the race . The K3 launch signals that Moonshot can deliver frontier-level open models at a fraction of the cost of US rivals – a key selling point for global businesses questioning the ROI of Anthropic and OpenAI pricing . The open-weight strategy also positions K3 as a direct competitor to DeepSeek's open models for self-hosted enterprise deployments .