Affiliate disclosure: This post contains affiliate links. If you buy through them, we may earn a small commission at no extra cost to you. Read our full disclosure.
Research-based analysis. We track launch coverage, benchmark data, and primary documentation rather than running our own lab tests. Every figure below is sourced and dated.
How we assess: this article is based on official documentation/manufacturer specs, pricing pages, and verified user/owner reports, not hands-on testing.
Moonshot AI just released the largest open-weight AI model anyone has ever shipped. Kimi K3 landed on July 16, 2026 with 2.8 trillion parameters, a 1 million token context window, and a benchmark run that put it at #1 on the Frontend Code Arena ahead of every closed model from OpenAI, Anthropic, and Google. The weights themselves arrive by July 27. That ten-day gap between announcement and download is exactly where hype tends to outrun facts, so this post separates what is confirmed from what is still marketing.
What is Kimi K3? Kimi K3 is Moonshot AI’s 2.8-trillion-parameter open-weight model, released July 16, 2026. It uses a mixture-of-experts design with 896 experts (16 active per token), handles text and images, offers a 1M-token context window, and costs $3 per million input tokens and $15 per million output. Downloadable weights are promised by July 27, 2026.
Key takeaways
- 2.8 trillion parameters total, with only 16 of 896 experts active per token (roughly 1.8% activation), per Tom’s Hardware.
- #1 on Frontend Code Arena at 1,679 points, ahead of Claude Fable 5, and an Artificial Analysis Elo of 1547, up 732 points from Kimi K2.6.
- $3 / $15 per million tokens in and out, with cache-hit input at $0.30, about 3x the price of Kimi K2.6.
- 1M-token context window with native image input, confirmed on OpenRouter’s model listing.
- Weights ship by July 27, 2026, so every “I ran it locally” claim you see before then deserves skepticism.
What Is Kimi K3 and Why Did It Break the AI News Cycle?
Kimi K3 is the new flagship model from Moonshot AI, the Beijing-based lab behind the Kimi chatbot line, and it broke the news cycle because it is the first open-weight model in the 3-trillion-parameter class. CNBC covered the launch on July 17 as a direct challenge to OpenAI and Anthropic. VentureBeat called it the largest open-source release ever. The claim that matters most for working developers: on coding and agentic benchmarks, Moonshot’s published numbers show K3 beating Claude Opus 4.8 and GPT-5.5, while trailing only the newest closed frontier models, Claude Fable 5 and GPT-5.6 Sol.
Context makes this bigger than one model. OpenAI shipped GPT-5.6 on July 9. xAI pushed out Grok 4.5 on July 8. Google delayed Gemini 3.5 Pro this same week after internal test failures, and Alphabet stock dropped 4% on the news. Into that window, a Chinese lab dropped a model with downloadable weights that undercuts closed-model pricing while matching last-generation flagships. If you want the pattern, read our Gemini 3.5 Pro confirmed-vs-hype breakdown. The open-versus-closed gap that looked like 12 months in 2024 now looks like weeks.
How Big Is Kimi K3, Really?
The honest answer: 2.8 trillion total parameters, but only about 50 billion do work on any given token. K3 is a mixture-of-experts (MoE) model with 896 expert networks, of which 16 activate per token, an activation rate near 1.8% according to Tom’s Hardware’s teardown of the technical report. That distinction matters because “2.8T” describes storage cost, not per-token compute. You still need serious hardware to host it. Moonshot recommends supernodes of 64 or more accelerators, and the weights ship in MXFP4 quantization with MXFP8 activations to keep memory manageable.
Two architecture choices stand out from the technical report. Kimi Delta Attention, a hybrid linear attention scheme, is how Moonshot claims a 1M-token context window without quadratic cost blowup. Attention Residuals modify how information flows between layers. Moonshot also built MiniTriton, an in-house compiler, and claims 2.5x better scaling efficiency than Kimi K2. One more detail with geopolitical weight: training reportedly ran on Nvidia H200s plus an unnamed second GPU vendor, which is how a Chinese lab works around U.S. export limits. None of these claims has independent verification yet; they come from Moonshot’s own report, and replication becomes possible only after the weights drop.
Can Kimi K3 Actually Beat Claude and GPT at Coding?
On one public leaderboard it already has: K3 sits at #1 on the Frontend Code Arena with 1,679 points, ahead of Claude Fable 5, which is a first for any open-weight model on that board. The Frontend Code Arena ranks models by human preference on real web-coding tasks, so it is hard to game with test-set memorization. Simon Willison’s July 16 write-up adds the broader picture: Artificial Analysis scores K3 at an overall Elo of 1547, a 732-point jump over Kimi K2.6, and K3 uses 21% fewer output tokens than K2.6 to get there.
Read the fine print before switching your stack, though. “Beats Claude at frontend code” is not “best model overall.” Aggregate rankings still place K3 fourth among frontier models, behind Claude Fable 5 and GPT-5.6 Sol and roughly even with or ahead of Claude Opus 4.8, depending on the eval. The model runs a single always-on reasoning level Moonshot calls “max,” which helps benchmarks and costs you tokens in production. For how the closed-model incumbents got here, our GPT-5.4 review and Claude Opus 4.7 review cover the previous generation’s baselines. Agentic coding harnesses like the ones we examined in our Devin review are where K3’s tool-use scores will face their real test.
How Much Does Kimi K3 Cost to Use?
API pricing is $3.00 per million input tokens and $15.00 per million output tokens, with cache-hit input discounted to $0.30 per million. Those numbers are live on OpenRouter and match Moonshot’s own listing. Two comparisons frame it. Against its predecessor: K2.6 cost $0.95 in and $4 out, so K3 is roughly 3x more expensive. Against the closed competition: $15 output undercuts flagship-tier pricing from OpenAI and Anthropic while the benchmark gap to those flagships is a few points, not a chasm.
The always-on reasoning changes the real math. Willison’s now-standard pelican-on-a-bicycle SVG test cost him 25 cents for a single prompt: 95 input tokens ballooned into 16,658 output tokens, 13,241 of them reasoning. You cannot turn that off. Budget for output-heavy bills if your workload involves long agentic chains, and use prompt caching aggressively since the $0.30 cache-hit rate is where the pricing gets attractive. Image input is cheap by comparison; Willison’s vision test cost 0.6 cents and produced what he called very good alt text.
What’s Confirmed and What’s Still Unverified?
Confirmed, from live product pages and multiple independent outlets: the July 16 release on Kimi.com, mobile apps, Kimi Code, and the API; 2.8T total parameters with the 896/16 MoE split; 1M context on OpenRouter; the $3/$15 pricing; the Frontend Code Arena #1 slot; native image input; and the public promise of open weights by July 27, 2026.
Unverified until weights land: every efficiency claim in the technical report, including the 2.5x scaling figure, the Kimi Delta Attention performance story, and the training-hardware details. Also treat third-party benchmark aggregates as provisional; early API endpoints sometimes differ from final released weights. Not yet known: the exact license text. “Open weight” is not one thing; K2 shipped under a modified MIT license, and until the K3 license posts, commercial teams should hold off on hard commitments. No English-language source we checked has published final license terms as of July 20.
How Was Kimi K3 Trained, and Why Does the Hardware Story Matter?
Moonshot’s technical report says K3 trained on Nvidia H200 accelerators plus hardware from a second, unnamed GPU vendor, and that detail carries more weight than any benchmark. U.S. export controls exist specifically to keep frontier-scale training out of Chinese labs, and a 2.8-trillion-parameter model is the largest possible counterexample. Tom’s Hardware framed the release exactly that way: China working around U.S. compute limits and shipping the biggest open-weight model ever anyway. Whatever the second vendor turns out to be, the practical takeaway is that compute scarcity did not stop this training run, and policymakers watching the export-control debate now have a concrete data point against the strategy.
The efficiency claims are the other half of the story. Moonshot says it built MiniTriton, an in-house compiler in the style of OpenAI’s Triton, to squeeze more out of whatever mixed fleet it was running, and it credits Kimi Delta Attention plus Attention Residuals for a claimed 2.5x scaling-efficiency gain over Kimi K2. The shipped weights use MXFP4 quantization with MXFP8 activations, aggressive low-precision formats that cut memory roughly in half versus the FP8 releases common a year ago. Every one of those numbers comes from Moonshot’s own report; none has outside replication yet. That is not a reason to dismiss them, since K2’s technical claims largely held up after release, but it is a reason to label them. The claims become checkable on July 27 when the weights land, and the speed with which independent labs confirm or puncture the 2.5x figure will tell you a lot about how much to trust the rest of the report.
What Does Kimi K3 Mean for the Open vs Closed AI Race?
It compresses the gap to almost nothing, at least on paper. The timeline of this single ten-day stretch makes the point: xAI shipped Grok 4.5 on July 8, OpenAI shipped GPT-5.6 on July 9, Google delayed Gemini 3.5 Pro after failed internal tests on July 17 and watched Alphabet drop 4%, and Moonshot released K3 into the middle of it on July 16 with an open-weights promise attached. Two years ago the standing assumption was that open models trailed closed flagships by about a year. K3 trails the very newest closed models by single-digit benchmark points in the same news cycle, and it beats the previous closed generation, Claude Opus 4.8 and GPT-5.5, on most published coding and agentic evals.
The competitive field around K3 sharpens the picture. Moonshot is not the only lab betting on open weights: Mira Murati’s Thinking Machines released Inkling, a 975-billion-parameter open-weight model, the same week, and DeepSeek is reportedly raising at a $74 billion valuation while pricing its API around 75% below Western competitors. On the infrastructure side, Fireworks AI just raised $1.5 billion at a $17.5 billion valuation specifically to serve open-model inference, which tells you investors expect self-hosted and third-party-hosted open models to take real workload share from closed APIs. For US and European buyers, the strategic effect is price pressure from two directions at once: open weights cap what closed labs can charge for last-generation performance, and Chinese API pricing caps what anyone can charge for mid-tier performance. The closed labs’ remaining moat is the frontier itself, plus trust, compliance, and data-residency guarantees that a Beijing-hosted API cannot offer Western enterprises. That moat is real, but it is narrower than it was in June.
How Can You Try Kimi K3 Right Now?
Four routes exist as of July 20, ordered by effort. Easiest: the hosted chatbot at Kimi.com and Moonshot’s mobile apps, which run K3 with free access tiers and give you a feel for its reasoning style and vision handling in minutes. Second: Kimi Code, Moonshot’s coding tool, if you want to test the agentic-coding claims in something closer to a real workflow. Third: the API, either direct from Moonshot or through OpenRouter, which is the sensible path for anyone already running multi-model setups since it needs no new account relationship with a Chinese provider and lets you A/B against your current model with a one-line change. Fourth, from July 27: the downloaded weights, for teams with data-center hardware or a rented cluster.
If you run the API test, structure it before you start. Pick ten to twenty real tasks from your own backlog rather than public benchmark prompts, because K3’s leaderboard results already price in benchmark-style questions. Turn on prompt caching from the first request; the gap between $3.00 cache-miss and $0.30 cache-hit input pricing is the difference between an expensive experiment and a cheap one for iterative coding sessions that resend large contexts. Log output token counts per task, not just quality scores, since the always-on reasoning mode is where budgets die quietly. And keep one eye on latency: reasoning-heavy responses that burn 13,000+ thinking tokens, like Willison’s pelican test did, take real wall-clock time, which matters if you are putting this behind a user-facing product rather than a batch pipeline.
What Are Kimi K3’s Biggest Limitations?
Four stand out from launch-week evidence. First, the always-on “max” reasoning is a cost and latency tax you cannot opt out of. Competing models ship adjustable reasoning effort; K3 gives you one gear. Willison’s single-prompt bill of 25 cents is trivial in isolation and brutal at scale: a thousand similar calls a day is $250, most of it reasoning tokens you never see. Second, the license is still unpublished. “Open weight” has ranged from genuinely permissive (K2’s modified MIT) to restrictive pseudo-open across the industry, and a commercial team that builds on K3 before reading the license text is taking a real legal risk for no reason when the answer arrives within a week.
Third, hosting is heavyweight even by open-model standards. MXFP4 quantization helps, but Moonshot’s own recommendation of 64+ accelerator supernodes puts casual self-hosting out of reach; the realistic paths are Moonshot’s API, aggregators, or US inference hosts once weights land. Fourth, everything impressive about K3 is currently unverifiable. The benchmark numbers come from Moonshot’s report and early API access, the efficiency claims have no replication, and the hosted endpoint could differ from the released weights in either direction. None of these is disqualifying. Together they say: treat K3 as a strong candidate pending verification, not a completed verdict, and let the July 27 weights release settle the open questions before you migrate anything that matters.
Should You Switch to Kimi K3?
Map it to your situation rather than the leaderboard. If you’re a frontend-heavy developer paying flagship API prices, run a two-week bake-off through OpenRouter now; the Arena result says K3 is at least competitive on your exact workload, and the cache-hit pricing rewards iterative coding loops. If you’re an enterprise with data-residency or IP concerns about Chinese-hosted APIs, wait for the July 27 weights and self-host; that is the entire point of open weights, and 64-accelerator supernodes are within reach for mid-size infra teams or via U.S. inference hosts like Fireworks AI. If you’re a solo builder on a budget, K3’s always-on reasoning makes it the wrong default for high-volume simple tasks; keep a cheap fast model for those and reserve K3 for hard problems. If you’re already deep in an editor stack like the ones in our Cursor review, swapping the underlying model is a config change, so testing costs you an afternoon.
How Does Kimi K3 Compare to GPT-5.6 and Claude?
Here is the frontier picture as of July 20, 2026, using published specs and pricing. Benchmark positions summarize launch-week coverage; treat single-digit gaps as noise.
| Model | Maker | Released | Open weights? | Context | API price (in/out per M) | Coding standing (July 2026) |
|---|---|---|---|---|---|---|
| Kimi K3 | Moonshot AI | Jul 16, 2026 | Yes, by Jul 27 | 1M | $3 / $15 | #1 Frontend Code Arena; ~4th overall |
| GPT-5.6 | OpenAI | Jul 9, 2026 | No | Not fully public | Tiered, flagship pricing | Top-2 overall (GPT-5.6 Sol) |
| Claude Fable 5 | Anthropic | 2026 | No | 1M-class | Flagship pricing | Top overall; #2 Frontend Arena |
| Claude Opus 4.8 | Anthropic | 2026 | No | 1M-class | Flagship pricing | K3 beats it on most coding/agentic evals |
| Grok 4.5 | xAI | Jul 8, 2026 | No | Not fully public | Mid-flagship | Competitive, below top-3 |
| Kimi K2.6 | Moonshot AI | 2025 | Yes | 256K | $0.95 / $4 | Superseded; 732 Elo behind K3 |
The strategic read: for the first time, an open-weight model is trading blows with closed flagships in the same release week, not a year later. Whether that holds after independent testing of the downloadable weights is the question the next ten days will answer. You can try the hosted model now at Kimi.com or through OpenRouter.
Popular AI gadgets & books on Amazon
Affiliate disclosure: As an Amazon Associate, AITrendyReview earns from qualifying purchases. Some links below are affiliate links, and we may earn a commission at no extra cost to you. This never changes a verdict.
Into AI hardware and reading too, not just software? A few of the most popular AI gadgets and books on Amazon right now:
- EMOPET AI Desk Robot — a ChatGPT-enabled desktop companion with voice commands and personality.
- Enzemit AI Translator Glasses — real-time translation across 138 languages, built into Bluetooth glasses.
- Best-selling books on AI — from beginner primers to prompt-engineering and machine-learning guides.
🤖 See more AI gadgets & books on Amazon →
FAQ: Kimi K3
When was Kimi K3 released?
Moonshot AI released Kimi K3 on July 16, 2026 across Kimi.com, its mobile apps, the Kimi Code tool, and its API, with aggregators like OpenRouter listing it the same week. Downloadable open weights are promised by July 27, 2026.
Is Kimi K3 free to use?
The hosted chatbot at Kimi.com has free access tiers, but the API is paid: $3 per million input tokens ($0.30 on cache hits) and $15 per million output tokens. Self-hosting becomes possible once weights ship, though you pay in hardware instead.
How many parameters does Kimi K3 have?
2.8 trillion total parameters in a mixture-of-experts design with 896 experts, only 16 of which activate per token. Per-token compute is closer to a ~50B dense model than to the headline number.
Is Kimi K3 better than GPT-5.6 or Claude?
It leads the Frontend Code Arena at 1,679 points, ahead of Claude Fable 5, and beats Claude Opus 4.8 and GPT-5.5 on most published coding and agentic benchmarks. Overall rankings still place it about fourth, behind Claude Fable 5 and GPT-5.6 Sol.
Can I run Kimi K3 locally?
Not yet, and for most people not ever on consumer hardware. Weights arrive by July 27, 2026 in MXFP4 quantization, and Moonshot recommends 64+ accelerator deployments. Realistic self-hosting means data-center GPUs or renting capacity from an inference provider.
Does Kimi K3 support images?
Yes. K3 accepts image input natively, and early testing (Simon Willison’s July 16 write-up) found its image description quality strong, at a cost of well under a cent per image at current API rates.
What license will the Kimi K3 weights use?
Unconfirmed as of July 20, 2026. Kimi K2 used a modified MIT license, but Moonshot has not published final K3 license terms. Commercial teams should wait for the actual text before building on it.
Why does Kimi K3 cost more than Kimi K2.6?
K3’s pricing ($3/$15) is roughly triple K2.6’s ($0.95/$4), and its always-on “max” reasoning produces long outputs. Moonshot is pricing it as a frontier product, betting the benchmark gains justify the premium.
Sources
- Tom’s Hardware: Moonshot releases 2.8-trillion-parameter Kimi K3 — Verified July 20, 2026
- Simon Willison: Kimi K3, and what we can still learn from the pelican benchmark — Verified July 20, 2026
- OpenRouter: Kimi K3 pricing and specs — Verified July 20, 2026
- CNBC: China’s Moonshot AI unveils Kimi K3 that rivals OpenAI, Anthropic — Verified July 20, 2026
- Codersera: Kimi K3 complete guide — Verified July 20, 2026