Kimi K3: Moonshot AI Drops a 2.8-Trillion-Parameter Open Model — and It Codes, Designs Chips, and Edits Video

#kimik3 #moonshotai #opensource #llm

TL;DR — Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter Mixture-of-Experts model with a 1-million-token context window and native vision. It's the largest open-weight model ever announced, competitive with GPT-5.5 and Claude Opus 4.8 on most benchmarks, and #1 (preliminary) on the Arena WebDev leaderboard. But it's also the most expensive Chinese model yet at $15/M output tokens, only runs in max-reasoning mode, and won't have downloadable weights until July 27. The launch, timed hours before Google's expected Gemini 3.5 Pro release, is the most aggressive product-timing move of 2026.

The 2.8T Elephant in the Room

Late on July 16, Chinese AI lab Moonshot AI published a blog post announcing Kimi K3 — a 2.8-trillion-parameter sparse Mixture-of-Experts model that instantly became the largest open-weight model in existence, roughly 1.75× the size of DeepSeek V4 Pro's 1.6T parameters. The launch was strategic to the minute: it landed hours before Google's expected Gemini 3.5 Pro debut on July 17, ensuring every Gemini benchmark published today will be compared against a model whose weights will be free within ten days.

(Source : Moonshot AI — Kimi K3: Open Frontier Intelligence)

Kimi K3 is not just bigger. It's built on two architectural novelties — Kimi Delta Attention (KDA) for efficient long-sequence scaling, and Attention Residuals (AttnRes) for selective depth-wise information retrieval — combined with a Stable LatentMoE router that activates 16 of 896 experts per token. Moonshot claims a ~2.5× improvement in scaling efficiency over its predecessor Kimi K2, though this is a vendor-reported figure, not independently reproduced.

(Source : Graphify — Kimi K3: Architecture, Benchmarks, Pricing, and Open Weights)

The model is available today via Kimi.com, Kimi Work, Kimi Code, and the Kimi API. But the "open" in "open 3T-class model" comes with an asterisk: the full weights won't be downloadable until July 27, 2026. Until then, Kimi K3 is a hosted model with an open-weight promise.

Benchmarks: Competitive, Not Dominant

Kimi K3's self-reported benchmarks place it in the tier below the frontier: it beats Claude Opus 4.8 and GPT-5.5 across most evaluations, but trails Claude Fable 5 and GPT-5.6 Sol. That's not a defeat — it's the most capable open-weight model ever measured, and it does it at roughly half the API cost of Opus 4.8.

(Source : Simon Willison — Kimi K3, and what we can still learn from the pelican benchmark)

The coding numbers tell the most interesting story. On Moonshot's table, Kimi K3 scores 67.5 on DeepSWE, 88.3 on Terminal-Bench 2.1, 81.2 on FrontierSWE, and 42.0 on SWE Marathon. It leads or ties on several coding benchmarks, though the footnotes reveal different harnesses were used across models (Kimi Code, Claude Code, Codex), making direct comparison noisy.

Where K3 genuinely shines is front-end development. On July 16, the live Arena WebDev leaderboard placed kimi-k3 at #1 with a score of 1,679 based on 1,757 votes — ahead of both Fable 5 and GPT-5.6 Sol. The result is marked preliminary with a ±17 interval, so it may shift, but it's the strongest independent signal so far.

Artificial Analysis adds a useful cost lens: Kimi K3 scores 57 on its Intelligence Index, outputs at 62 tokens/sec, and consumed ~130 million output tokens during evaluation — more than double the comparison average. The total evaluation cost: $2,690.80. In plain English: K3 is smart, but it's verbose, not particularly fast, and not cheap at uncached API rates.

(Source : Artificial Analysis — Kimi K3)

What 2.8T Parameters Actually Buys You

The most arresting part of Moonshot's launch isn't the benchmark table — it's the case studies. These are vendor-reported and should be read with that caveat, but they describe capabilities that would have been science fiction eighteen months ago:

Kernel Optimization. Kimi K3 was given 24 hours in a sandbox to profile, rewrite, and benchmark GPU kernels across H200 and alternative GPGPU hardware. It performed competitively with Claude Fable 5 (which may have used fallback behavior) and "substantially outperformed" Opus 4.8, GPT-5.6 Sol, and GPT-5.5. Moonshot says an early version of K3 handled the majority of the team's kernel optimization work during late-stage development.

GPU Compiler from Scratch. K3 built "MiniTriton" — a compact Triton-like compiler with its own tile-level IR, MLIR integration, optimization passes, and a PTX code-generation pipeline. On supported roofline benchmarks, MiniTriton matched or beat Triton and torch.compile. It sustained end-to-end nanoGPT training with stable convergence. This is not isolated kernel tweaking; it's a coherent end-to-end compiler built by a model.

Chip Design. In a 48-hour autonomous run, K3 designed a chip for a nano model built on its own architecture using open-source EDA tools on the Nangate 45nm library. The result: a 4 mm² chip closing timing at 100 MHz, sustaining 8,700+ tokens/sec decode throughput, packing 1.46M standard cells, 0.277 MB of SRAM, and an INT4 MAC array. A model designing silicon for a model.

Research Automation. K3 reproduced the I-Love-Q universal relations from computational astrophysics in about two hours — work Moonshot estimates would take an experienced researcher one to two weeks. It reviewed 20+ papers, implemented the full numerical pipeline, evaluated 300+ equations of state, identified formula inconsistencies, and generated 3,000+ lines of Python.

(Source : Moonshot AI — Kimi K3 Tech Blog)

These demos are spectacular. They are also vendor-curated. The honest read is that Kimi K3 has genuine long-horizon agentic capability that extends well beyond chatbot benchmarks, but we don't yet know how reliably it performs these tasks or how much human intervention the demos required.

Pricing: The Most Expensive Chinese Model Yet

Kimi K3's API pricing is $3 per million input tokens and $15 per million output tokens — roughly on par with Anthropic's Claude Sonnet series, and a massive jump from Kimi K2.6's $0.95/$4 pricing. It's the most expensive model ever released by a Chinese AI lab.

Simon Willison ran his famous "pelican riding a bicycle" SVG test through K3. The result: 95 input tokens, 16,658 output tokens (13,241 of which were reasoning tokens), costing 25 cents for a single SVG generation. For context, the same test on GPT-5.6 Sol costs significantly less, and on smaller open models it's effectively free.

(Source : Simon Willison — Kimi K3 pelican test)

The pricing makes sense given the architecture. With only max-reasoning effort available at launch (low and high modes are promised later), every query burns through reasoning tokens. Moonshot says this is temporary, but for now, K3 is both the most capable and the most expensive open-model API on the market.

The Strategic Timing: Why July 16 Matters

Kimi K3 didn't land on a random Tuesday. It dropped the night before Google's expected Gemini 3.5 Pro launch — an event already framed as a "make-or-break" moment after Google reportedly scrapped the original base model and restarted pretraining six weeks ago. The BuildFastWithAI roundup captured the dynamic perfectly: "the split-screen of today is the whole year in one image. The West's biggest model launch, the East's biggest AI policy speech, and a Chinese open model dropped between them, all within 24 hours."

(Source : BuildFastWithAI — AI News Today July 17 2026)

The same day, Xi Jinping keynoted the World AI Conference in Shanghai for the first time, proposing a World AI Cooperation Organization headquartered in the city. The Kimi K3 launch the night before was not a coincidence of calendars — it was a demonstration that Chinese labs can compete on capability at the exact moment China volunteers to hold the pen on global AI governance.

For developers, the practical question is simpler: when the K3 weights land on July 27, what exactly justifies the premium on closed models? A 2.8T open-weight model that can design chips and build compilers, available for anyone to run (hardware permitting), resets the value proposition for every API-priced frontier model above it.

FAQ

Q: Is Kimi K3 actually open source right now?
A: No. The weights are promised by July 27, 2026. Until then, K3 is a hosted model accessible via Moonshot's products and API. "Open source" also means different things — weights, training code, data, and license terms are separate questions.

Q: How does K3 compare to GPT-5.6 Sol and Claude Fable 5?
A: It trails both on most benchmarks, but is competitive with GPT-5.5 and Opus 4.8. It currently leads the Arena WebDev leaderboard (preliminary). For coding agents, it's in the conversation; for general reasoning, it's one tier below the frontier.

Q: Can I run K3 locally?
A: Not realistically on consumer hardware. At 2.8T parameters with 16/896 expert sparsity, even quantized inference will require enterprise-grade GPU clusters. The practical path for most developers will be the hosted API or cloud inference partners once weights are released.

Q: Why is it so expensive compared to other Chinese models?
A: Max-reasoning mode is the only option at launch, burning reasoning tokens on every query. Lower-effort modes are promised in future updates. The base pricing ($3/$15) also reflects the massive scale — this is a frontier-class model priced accordingly.

Q: What happens on July 27?
A: Moonshot promises to release the full model weights, along with a technical report covering architecture, training, and evaluations in detail. The open-weight community will then port it to inference engines like vLLM, llama.cpp, and SGLang.