Kimi K3: Full Specifications, Benchmark Results, API Pricing, Open-Weight Status and Verdict
Moonshot AI has unveiled Kimi K3, a 2.8 trillion-parameter model with a 1 million-token context window, native image understanding and a strong focus on long-horizon coding, research and agentic work. Early results place it near leading proprietary systems, but the promised model weights are not yet publicly downloadable.
Kimi K3 Is a Serious Frontier Challenger, but Its Open-Weight Claim Is Not Yet Complete
Kimi K3 is one of the most important AI model announcements of 2026. Independent evaluations indicate that it can compete near the top in selected coding, web-development and professional-task benchmarks. Its scale and context window also make it technically unusual.
The model should not yet be called a fully released open-source system. Moonshot says the weights will arrive by July 27, 2026. Until the files, license and technical report are published, the most accurate description is a planned or announced open-weight release.
What Kimi K3 Changes in the AI Model Race
Kimi K3 changes the conversation because it combines extreme model scale, a very large context window and competitive early benchmark performance in a system that Moonshot intends to release with downloadable weights.
The announcement directly challenges the assumption that the strongest open-weight systems must remain far behind closed frontier models. It also weakens the idea that Chinese laboratories consistently trail major US developers by a predictable margin.
That does not mean Moonshot has overtaken Anthropic or OpenAI overall. Moonshot itself says K3 still trails Claude Fable 5 and GPT-5.6 Sol in overall performance. The more defensible conclusion is that the gap is uneven, task-dependent and becoming harder to dismiss.
Kimi K3 is not proven to be the best AI model overall. It is a high-scale model with enough independent benchmark strength to show that planned open-weight systems are moving much closer to the proprietary frontier.
Kimi K3 Complete Specifications
| Developer | Moonshot AI |
|---|---|
| Model | Kimi K3 |
| Total parameters | 2.8 trillion |
| Architecture | Sparse Mixture of Experts |
| Expert routing | Moonshot says 16 of 896 experts are effectively activated during processing |
| Exact active parameters | Not yet publicly disclosed |
| Context window | Up to 1 million tokens |
| Input support | Text and images |
| Output | Text |
| Primary strengths | Coding, agentic workflows, research, visual analysis and professional knowledge work |
| Current access | Kimi, Kimi Work, Kimi Code and Kimi API |
| Weights release | Promised by July 27, 2026 |
| Technical report | Pending at publication |
| Recommended self-hosting scale | At least 64 accelerators, according to Moonshot |
The 2.8 trillion figure refers to the total parameter pool, not the amount of compute used for every token. The undisclosed active-parameter count is more useful for estimating practical inference requirements.
Why 2.8 Trillion Parameters and a 1M-Token Context Matter
Mixture of Experts makes the scale possible
Kimi K3 uses a sparse Mixture of Experts design. Rather than activating the entire model for every token, a routing system selects a smaller set of specialist components. Moonshot says 16 of 896 experts are effectively activated during processing.
This architecture allows K3 to draw from an enormous parameter pool without paying the full computational cost of a dense 2.8 trillion-parameter model on every step.
Input
The model receives text, an image or task context.
Routing
The router identifies the expert components most relevant to the token.
Expert compute
A small subset of the total expert pool performs the computation.
Combined output
The selected expert representations contribute to the response.
What a 1 million-token context can support
Large software repositories
K3 can receive more source files, documentation, requirements and task history within one working session.
Multi-document research
The context capacity can support comparisons across reports, papers, contracts and supporting evidence.
Long-running autonomous agents
More interaction history can remain available during research, coding, testing and revision.
Perfect recall across 1 million tokens
Accepting a long prompt does not prove that the model will retrieve, connect and apply every important detail accurately.
Model scale and context length are capacity indicators, not direct measures of intelligence. Training quality, routing, post-training, tool use, inference design and evaluation conditions remain critical.
Kimi K3 Benchmarks: Independent Results vs Moonshot Claims
The benchmark picture is encouraging, but the evidence should be divided into independent evaluations and company-reported demonstrations.
Independent evaluation snapshot
| Evaluator | Reported result | What it suggests | Important caution |
|---|---|---|---|
| Artificial Analysis | 57 Intelligence Index; fourth among 189 tracked models | Near-frontier composite capability | Slower output and unusually high output-token usage |
| Arena WebDev | Preliminary score of 1,679; first at the checked date | Strong front-end web-development preference results | Preliminary and limited to web development |
| Vals AI | Reported second overall in a professional-task evaluation | Competitive domain-focused knowledge work | Not a universal ranking across every task |
Artificial Analysis also found important trade-offs
The same independent evaluation that placed K3 near the frontier also described it as slower than average, highly verbose and not automatically cheap against every similarly capable model.
Moonshot's GPU kernel result
Moonshot says K3 performed competitively with Claude Fable 5 and substantially outperformed Claude Opus 4.8, GPT-5.6 Sol and GPT-5.5 in a specialized GPU kernel-optimization test.
This is a narrow engineering evaluation reported by the model developer. It should not be converted into the broader claim that K3 is universally better than those systems.
- Models may use different agent harnesses and tools.
- Reasoning effort, time limits and fallback behavior may differ.
- Developer-selected tests can favor a model's intended strengths.
- Preliminary rankings can change as more evaluations are added.
- A category win does not establish overall model superiority.
Coding, Agentic Work and Engineering Demonstrations
Moonshot is positioning K3 as a long-horizon worker rather than only a chatbot. The company says it can inspect repositories, use terminal tools, review visual outputs, test results and continue revising its work with limited supervision.
MiniTriton compiler
Moonshot reports that K3 created a compact GPU programming system with an intermediate representation, compiler optimization passes and a PTX code-generation pipeline.
48-hour chip-design experiment
K3 reportedly designed, optimized and verified a small simulated chip intended to serve a model based on its own architecture. This was a proof of concept, not manufactured commercial silicon.
Computational research workflow
Moonshot says K3 reviewed more than 20 papers, implemented a numerical pipeline, evaluated hundreds of equations of state and generated an interactive research dashboard.
Visual software development
K3 is designed to use screenshots and visual feedback while refining interfaces, websites, games and other software outputs.
These demonstrations matter because they test sustained execution across planning, implementation, testing and revision. They still need independent reproduction using the released weights, prompts, tools and evaluation environment.
Is Kimi K3 Really Open Source?
Not yet in the practical sense that most developers expect.
Moonshot calls K3 an open model and says the full weights will be released by July 27, 2026. At publication, the model was accessible through Moonshot's products and API, but independent users could not yet download and host the full weights.
Moonshot hosts the model and controls the serving infrastructure.
Users can independently host the trained weights under the published license.
Training details, data documentation, safety results and reproducibility materials are sufficiently disclosed.
Why the final license matters
Publishing weights does not automatically make a model fully open source. The license may restrict commercial use, redistribution, modification or specific applications. The final classification should wait for the actual files and terms.
Kimi K3 is an API-accessible proprietary model with an announced open-weight release. It should not be described as fully released open source until the weights and license are public.
Kimi K3 API Pricing and Deployment Requirements
Kimi K3 is available through Moonshot's API under the model identifier
kimi-k3.
Moonshot reports a cache-hit rate above 90 percent for coding workloads on its infrastructure. Actual cost depends on prompt reuse, context length, output volume, reasoning effort and the number of agent steps.
Why the listed token price can be misleading
K3's output price may appear competitive, but independent testing found that the model generated unusually large outputs. A verbose model can cost more per completed task even when its listed token rate is lower.
Can Kimi K3 run on a personal computer?
No realistic full deployment fits an ordinary consumer computer. Moonshot recommends supernode configurations containing at least 64 accelerators. The planned open-weight release may improve control and research access, but it will not make the full model cheap or simple to operate.
Market Impact and the China-US AI Competition
Reuters reported that K3's announcement was followed by sharp declines in the shares of Chinese AI competitors Zhipu and MiniMax. Shortly before the market close, Zhipu was down about 27.7 percent and MiniMax was down about 16.5 percent.
The timing suggests that investors viewed K3 as a threat to the scarcity value of competing proprietary-model portfolios. A capable open-weight release can pressure API pricing, subscriptions, enterprise margins and developer loyalty.
Pricing pressure
Competitors may need to lower API prices or justify a premium.
Lower model scarcity
Frontier-like capability becomes less exclusive as alternatives grow.
Private deployment options
Enterprises may value control over data, customization and hosting.
Faster release cycles
Labs may need to update models more frequently to retain attention.
A same-day stock decline does not prove that K3 alone caused the entire move. Valuation concerns, profit-taking, broader technology sentiment and short-term trading may also contribute.
What K3 actually proves about global competition
K3 does not establish that Chinese laboratories lead the United States across infrastructure, chips, enterprise adoption, safety engineering, distribution or overall model quality.
It does show that Chinese labs can produce models that compete near the frontier in important categories while using aggressive pricing and planned open-weight distribution as strategic pressure.
Who Should Use Kimi K3 and Who Should Wait?
- Long-horizon coding and repository analysis
- Large-context document and research workflows
- Visual feedback during software development
- A competitive API alternative to premium closed models
- Early access to a planned high-scale open-weight ecosystem
- Publicly downloadable weights today
- A fully documented open-source license
- Low-latency and concise responses by default
- Cheap self-hosting on ordinary hardware
- A mature safety report and complete technical documentation
Best fit: advanced developers and enterprise evaluators
K3 is most relevant to teams evaluating coding agents, large-document workflows, research automation and high-context professional tasks. Its API allows experimentation before the promised weights become available.
Weakest fit: users buying only by the 2.8T headline
Parameter count alone is not a reason to adopt a model. Teams should compare task success rate, latency, output length, tool reliability, total cost and governance requirements against alternative systems.
Final Verdict: A Major Model, but Not Yet a Completed Open Release
Kimi K3 is one of the strongest signals yet that the open-weight and proprietary frontiers are converging. Its 2.8 trillion-parameter sparse architecture, 1 million-token context window and competitive independent results make it a serious technical and strategic release.
The strongest evidence is not Moonshot's parameter headline. It is the combination of strong independent benchmark placement, credible coding performance and the market pressure created by the planned weight release.
The model still carries substantial uncertainty. The weights, final license and full technical report are pending. Independent testing also points to slower output and unusually high token usage, while the hardware requirements put full self-hosting beyond most organizations.
Treat Kimi K3 as a serious near-frontier model and an important planned open-weight release, not as proof that open AI has already defeated every closed system. Its real credibility test begins when Moonshot publishes the weights, license and technical report.
Frequently Asked Questions
What is Kimi K3?
Kimi K3 is Moonshot AI's 2.8 trillion-parameter multimodal model for coding, reasoning, research, visual analysis and professional knowledge work.
Is Kimi K3 open source?
Not yet. K3 is currently available through Moonshot's products and API. Moonshot says the full weights will be released by July 27, 2026, but the final license and technical report were still pending at publication.
Is Kimi K3 the world's largest open model?
Moonshot describes K3 as the world's first open model in the 3-trillion-parameter class. The claim should remain qualified until the promised weights and license are publicly available.
Does Kimi K3 activate all 2.8 trillion parameters?
No. K3 uses a sparse Mixture of Experts architecture. Moonshot says 16 of 896 experts are effectively activated during processing. The exact active-parameter count has not yet been disclosed.
How large is the Kimi K3 context window?
Kimi K3 supports a context window of up to 1 million tokens.
Is Kimi K3 better than GPT-5.6 Sol?
K3 performs better in some reported tests, but Moonshot says it still trails GPT-5.6 Sol overall. Results depend on the task, tools, reasoning effort and evaluation environment.
Is Kimi K3 better than Claude Fable 5?
K3 led Arena's preliminary WebDev ranking at the checked date, but Moonshot says Claude Fable 5 remains stronger overall. No single benchmark establishes universal superiority.
How much does the Kimi K3 API cost?
Moonshot lists K3 at $0.30 per million cache-hit input tokens, $3.00 per million cache-miss input tokens and $15.00 per million output tokens.
Can Kimi K3 run on a personal computer?
The full model is far beyond ordinary consumer hardware. Moonshot recommends deployment configurations with at least 64 accelerators.
When will the Kimi K3 weights be released?
Moonshot says the full model weights will be released by July 27, 2026.
Sources and Verification
- Moonshot AI - Kimi K3 Technical Blog
- Reuters - Moonshot Unveils Kimi K3
- Artificial Analysis - Kimi K3 Intelligence, Speed and Pricing
- Arena - WebDev Leaderboard
- Vals AI - Kimi K3 Model Evaluation
- Kimi API Platform
Specifications, benchmark positions, pricing and availability may change. Company demonstrations are labeled separately from independent evaluations. Verify the release status of the weights and license after July 27, 2026.