Moonshot launches Kimi K3: the world's largest open-weight model at 2.8 trillion parameters — closing in on Claude Fable 5 and GPT-5.6 Sol

On July 16, 2026, Chinese AI lab Moonshot AI (backed by Alibaba) announced Kimi K3 — a 2.8-trillion-parameter model the company calls the world's largest open-weight model. The full weights were publicly released on Hugging Face on July 27, 2026, under Moonshot's own "Kimi K3 License". The headline: independent evaluations place K3 just behind today's two strongest closed models, Claude Fable 5 and GPT-5.6 Sol — at a fraction of their API prices.

What is Kimi K3?

Per Kimi's official blog, K3 is Moonshot's most capable model to date: 2.8 trillion parameters, built on two architectural updates — Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) — with native vision (not a bolted-on module) and a 1-million-token context window. Moonshot calls it "the world's first open 3T-class model," aimed at long-horizon coding, knowledge work and deep reasoning.

Architecturally, K3 pushes Mixture-of-Experts sparsity further: each query activates just 16 of 896 experts. Combined with refined training recipes, Moonshot says K3 converts compute into intelligence about 2.5× more efficiently than Kimi K2. The model reasons by default — at announcement (July 16, 2026) the API exposed a single reasoning_effort level, max, with Moonshot saying lighter modes would follow.

Benchmarks: third in the world, ahead of Opus 4.8

The numbers drawing attention come from independent evaluations. Per Artificial Analysis figures cited by Moonshot in the K3 blog: on GDPval-AA v2 — a leaderboard testing models on real work tasks across 44 occupations in 9 industries — K3 scored 1,668, behind only Claude Fable 5 Max (1,760) and GPT-5.6 Sol Max (1,748), and ahead of Claude Opus 4.8 Max (1,600). On AA-Briefcase (Artificial Analysis' private benchmark for long-horizon agentic knowledge work), K3 scored 1,548 — second behind Fable 5 Max (1,583), ahead of GPT-5.6 Sol Max (1,495).

On front-end coding, Arena.ai's Code Arena leaderboard ranks K3 above Fable 5 — a 17-place jump over its predecessor Kimi K2.6. Arena CEO Anastasios Angelopoulos commented on X that it may be "the single biggest release of the year." For long information-seeking tasks, Moonshot self-reports a 91.2 BrowseComp score running as a single agent — a result that uses a context-compaction strategy triggered at 300K tokens (as applied in the Claude model cards); evaluated with the full 1M-token window and no context management, the score is 90.4. Both are company-reported figures awaiting independent verification.

Moonshot itself concedes in its blog that K3's overall performance still trails the two strongest proprietary models, Claude Fable 5 and GPT-5.6 Sol — while delivering frontier-level results and outperforming the other models it tested. The context matters: Fable 5 shipped only last month, and GPT-5.6 debuted a week before K3 — the open-vs-closed gap is now measured in weeks, not quarters.

Pricing: a third of the cost, full 1M-token window at flat rates

Kimi K3 is priced at $3 per 1 million input tokens ($0.30 on cache hits) and $15 per 1 million output tokens (reasoning tokens included) — and those rates apply across the entire 1,048,576-token window, with no long-context surcharge. For comparison, here are the official Anthropic and OpenAI price lists (standard rates, excluding batch discounts):

Quick math: K3's input price is 30% of Fable 5's and 60% of Opus 4.8 / GPT-5.6 Sol's; its output price is 30% of Fable 5's and 50% of GPT-5.6 Sol's. For reasoning-heavy workloads with long outputs, the cost gap compounds fast.

What can K3 do? Vision-in-the-loop and long-horizon coding

Moonshot positions K3 for extended engineering sessions: reading large codebases, orchestrating terminal tools, and correcting course on its own when an approach fails. The highlighted differentiator is "vision-in-the-loop" — the model screenshots its rendered output, looks at it, then goes back to fix the code, iterating until it works. Published demos include a browser-based 3D open-world game (Three.js + WebGPU), a simulation of the Long March 10 rocket's launch and return, and a Game Boy Advance emulator. In another case study, K3 ran autonomously for 48 hours to design and verify a chip for a nano model on the Nangate 45nm library using open-source EDA tools.

A second "DeepSeek moment"?

The last time a Chinese lab shipped a cheap model that matched closed U.S. models was DeepSeek R1 in January 2025 — an event that wiped roughly $1 trillion off leading U.S. tech stocks. Former White House AI policy adviser Sriram Krishnan called K3's debut a big moment with industry-wide implications. On the flip side, SiliconANGLE notes Anthropic previously accused Moonshot (along with DeepSeek and MiniMax) of violating its terms by distilling Claude's capabilities to train their own models — expect the debates around distillation and chip export controls to reignite.

As for Moonshot itself: per the Financial Times (via TechCrunch), the company is raising new capital at a $31.5 billion valuation — up sharply from the $2 billion round at a $20 billion valuation it closed in May 2026.

What it means for Vietnamese businesses

Kimi K3 reinforces the trend we analyzed in our DeepSeek V4 and AI sovereignty pieces: open models keep closing in on closed ones, at a fraction of the price. Three practical takeaways:

• Use the API today: for workloads without sensitive data, $3/$15 with a 1M-token context is highly competitive for long-document processing and large codebases.
• Self-host with the released weights: the weights shipped on July 27, 2026, so businesses can run K3 on their own infrastructure — data never leaves the organization. Caveat: 2.8 trillion parameters is a very large model; self-hosting requires a serious GPU cluster, so it suits organizations with existing infrastructure or private cloud. Our recommended framework remains data classification — routine tasks via API, core data on mid-sized open models on-premise — as laid out in our internal AI roadmap.
• Read the license before commercial use: "open-weight" does not automatically mean anything-goes. K3 ships under the "Kimi K3 License" — not the "Modified MIT" license used for K2. It permits use, modification, distribution and fine-tuning, but a Model-as-a-Service business with more than USD 20 million in revenue over 12 months must sign a separate agreement with Moonshot, and commercial products with more than 100 million monthly active users or more than USD 20 million in monthly revenue must prominently display "Kimi K3". Purely internal use (no software or outputs provided to third parties) is exempt from those two conditions.

Frequently asked questions

Note: This article aggregates public sources as of July 17, 2026; weight-release and license status updated on September 24, 2026 from Moonshot's Hugging Face page; benchmark and pricing figures are as published by Moonshot, Anthropic, OpenAI or the cited sources at the time of writing and may change. For reference only — not technical or investment advice.