Was ist Kimi K3 – und ist es GPT-5.6 und Fable 5 wirklich so nah?

What Is Kimi K3?

Kimi K3 is Moonshot AI’s new flagship multimodal reasoning model. It was built for work that may involve hours of coding, research, tool use, and revision—not just answering a single chat prompt.

The model went live on July 16, 2026, across Kimi.com, Kimi Work, Kimi Code, and the Kimi API. It has 2.8 trillion total parameters, a one-million-token context window, and native support for text, image, and video input. Its output is still text.

Under the hood, K3 uses a Mixture-of-Experts architecture with 896 experts, 16 of which are activated at a time. Moonshot also introduced Kimi Delta Attention, Attention Residuals, and Stable LatentMoE to improve how information moves through long sequences and deep model layers. Moonshot claims these changes produce roughly 2.5 times the scaling efficiency of Kimi K2, although that figure comes from the developer rather than an independent test. Moonshot’s Kimi K3 technical blog provides the current architecture details; a fuller technical report is still pending.

That makes K3 one of the largest AI models announced to date. Size, however, is not the useful part by itself. The real question is whether Moonshot has turned those parameters into better task completion.

The release also removes one uncertainty from the original launch coverage: the technical report is now public. It does not remove the need for independent evaluation, and it does not turn a 2.8T model into a low-friction self-hosting project.

Is Kimi K3 Really Close to GPT-5.6 and Claude Fable 5?

In overall intelligence, not quite. In several agentic and long-horizon tasks, yes.

Moonshot’s own announcement is more restrained than many of the headlines it produced. The company says K3’s overall performance still trails Claude Fable 5 and GPT-5.6 Sol. It nevertheless reports frontier-level results across coding, knowledge work, and reasoning.

One example is GPU kernel optimization. Moonshot placed each model in an identical sandbox and gave it up to 24 hours to optimize four GPU tasks. K3 performed competitively with Fable 5 and beat GPT-5.6 Sol, GPT-5.5, and Opus 4.8 in this test. That is a substantial result for low-level performance engineering. It is not evidence that K3 is universally smarter.

Independent testing gives us a broader picture. Artificial Analysis scored Kimi K3 at 57 on its Intelligence Index, placing it around GPT-5.5 and Claude Opus 4.8 but behind Fable 5 and GPT-5.6 Sol. Its strongest results appeared in agentic execution and automation rather than every category of intelligence. Artificial Analysis published its K3 findings on July 16.

The last row is the one that stopped me. K3 became more capable, but Artificial Analysis also measured more hallucination. It answered more questions correctly while producing a higher proportion of unsupported claims.

That trade-off matters in production. A model that completes a long agent workflow but quietly introduces an incorrect assumption can cost more to verify than a slower model with steadier factual behavior.

Why One Leaderboard Cannot Settle the Comparison

K3 has already reached the top of at least one frontend-generation leaderboard. It also performs well in GPU optimization, SaaS automation, and long-horizon knowledge work. Those results do not measure the same capability.

A frontend arena measures how users prefer generated interfaces. AutomationBench measures whether an agent can operate software workflows. GPU kernel optimization tests low-level systems engineering. None can independently answer whether K3 is better at research, factual recall, debugging, presentation design, or general reasoning.

The defensible conclusion is narrower: Kimi K3 has entered the frontier group for agentic work, but GPT-5.6 Sol and Fable 5 still lead the broader comparison. Developers who want the current closed-model ceiling can examine the GPT-5.6 Sol API, while Claude Opus 4.8 remains a more established option for complex coding and research workflows.

The Open-Weight Question Is Settled; the Deployment Question Is Not

The launch-week question was whether Moonshot would actually publish the weights. It has. The official moonshotai/Kimi-K3 repository now contains the full checkpoint, model card, technical report, inference code, and Kimi K3 License.

The precise wording still matters. Kimi K3 is open-weight, not an unqualified “fully open-source” release. Its custom license permits broad use, modification, fine-tuning, deployment, and redistribution, but adds conditions for large Model-as-a-Service businesses and very large commercial products. The training data is not public.

Availability is also not usability. The repository is about 1.56 TB, and Moonshot recommends supernode deployments with 64 or more accelerators. That makes K3 relevant to cloud and enterprise infrastructure teams, not a routine workstation download.

The release is still important. Organizations now have a real path to inspect the checkpoint, customize inference, fine-tune derivatives, and operate the model without depending entirely on one hosted endpoint. The price is hardware, serving engineering, and license review.

What Changed From Kimi K2.7 to Kimi K3?

K3 is not simply Kimi K2.7 with a larger version number.

Kimi K2.7 Code is a coding-focused agentic model based on K2.6. Its model card emphasizes real-world software engineering, long coding sessions, tool use, and lower reasoning-token consumption. K3 expands that role into a general flagship covering coding, visual reasoning, research, and end-to-end knowledge work.

The context window grows from 256K to one million tokens. Total parameters rise from one trillion to 2.8 trillion, while the expert count grows from 384 to 896. The price rises with it.

K3's value proposition is therefore not “frontier intelligence for almost nothing.” It is frontier-adjacent agentic performance with two deployment paths: a metered API for immediate use and released weights for teams prepared to own the infrastructure.

The official K2.7 Code model card contains its architecture, benchmark methodology, and deployment guidance.

How Kimi K3 Handles a 1M-Token Context

A one-million-token context window allows K3 to accept large codebases, collections of technical documents, long agent histories, and intermediate tool results in one request.

That is particularly relevant for long-running agents. A coding model may need to retain an initial specification, repository structure, terminal output, test failures, earlier edits, and reviewer feedback across dozens of steps. Losing one of those constraints halfway through the task is a common reason an agent appears productive but fails to finish.

K3 also applies automatic context caching. Developers do not need to create a cache ID or set a separate time-to-live value. If a long prefix remains unchanged between requests, the system automatically attempts a cache hit. Cached input costs $0.30 per million tokens instead of $3.00.

But context capacity is not the same as perfect context use. Feeding a model one million tokens does not guarantee that it will assign the correct importance to every line. Longer inputs can also increase cost, processing time, and distraction from irrelevant material.

The sensible approach is still to retrieve the most relevant files first, keep stable reference content at the beginning for caching, and avoid sending the entire knowledge base merely because the limit allows it. Kimi’s API guide explains its one-million-token context and automatic caching behavior.

Kimi K3 API Pricing Is No Longer Budget Pricing

Moonshot’s official Kimi K3 API uses flat pay-as-you-go pricing. The rate does not change when a request crosses a larger context tier.

For a shorter agent task using 100,000 uncached input tokens and producing 20,000 output tokens, the estimated model cost is:

(0.1 × $3) + (0.02 × $15) = $0.60

Now consider a repeated long-context workflow with 800,000 cached tokens, 50,000 new input tokens, and 50,000 output tokens:

(0.8 × $0.30) + (0.05 × $3) + (0.05 × $15) = $1.14

Caching keeps the second example manageable, but K3 is clearly not following the old “Chinese model equals extremely cheap API” formula.

Artificial Analysis measured an average Intelligence Index task cost of $0.94 for K3, close to GPT-5.6 Sol at $1.04 and roughly half of Opus 4.8 at $1.80. GLM-5.2 completed the same evaluation workload for considerably less.

The comparison also changes depending on where the models are accessed. The Kimi K3 API on GPT Proto currently costs $2.70 per million input tokens and $13.50 per million output tokens, 10% below Moonshot’s $3/$15 list price. GPT Proto also offers GPT-5.6 Sol and GLM-5.2 through the same API key and shared balance, making task-level comparisons easier without separate provider accounts.

K3's value proposition is therefore not “frontier intelligence for almost nothing.” It is frontier-adjacent agentic performance with two deployment paths: a metered API for immediate use and released weights for teams prepared to own the infrastructure.

Kimi K3 API vs Open Weights: Which Should You Use?

 Kimi K3 now gives developers a real deployment choice. The model capability is the attraction on both paths; the operating model is completely different.

Use the API first if you are still proving product-market fit, traffic is variable, or the team does not already operate large-model infrastructure. The GPT Proto Kimi K3 API also makes it easier to compare K3 with GPT-5.6 Sol, GLM-5.2, and other models behind one key and balance.

Use the weights when private deployment, fine-tuning, custom inference, or provider independence has enough business value to justify the cluster. Before commercial deployment, review the Kimi K3 License rather than assuming “open” means MIT.

Kimi K3 vs GLM-5.2, GPT-5.6 and Opus 4.8

The benchmark winner is not automatically the right production model. The useful choice depends on what can fail, how long the workflow runs, and whether owning the weights matters.

K3 has the clearest case when three conditions appear together: a long-running task, multimodal inputs, and a reason to control the model deployment. Without those requirements, its scale can become an expensive answer to a smaller problem.

GPT-5.6 Sol remains the stronger choice when broad intelligence and reliability come first. Opus 4.8 makes sense for teams with mature Claude-based coding or research workflows. GLM-5.2 is the harder competitor on economics because it already provides a one-million-token context and open weights at a lower task cost.

My current rule of thumb is simple: choose K3 when open weights and long multimodal agents matter. Choose GPT-5.6 or Opus when reliability matters more than model ownership. Choose GLM-5.2 when cost is the constraint you cannot negotiate.

Can You Use Kimi K3 Through GPT Proto Yet?

Yes. Kimi K3 is now available through the GPT Proto Kimi K3 API.

The current GPT Proto rate is $2.70 per million input tokens and $13.50 per million output tokens, 10% below Moonshot’s current $3/$15 list price. Use the model string kimi-k3 with your GPT Proto API key to access K3’s long-context coding, multimodal analysis, tool calling, and structured-output capabilities.

The same key and shared balance also cover GPT-5.6 Sol, GLM-5.2, Claude Opus 4.8, and 200+ text, image, video, and audio models. This makes it practical to test K3 against alternatives on the same repository or agent workflow before changing production traffic.

You can try Kimi K3 on GPT Proto or browse the full GPT Proto AI model gallery. The model's weights and technical report are now public, but the article should still be revisited as independent deployment tests, quantizations, throughput data, and framework support mature.

Final Verdict: Close, but the Benchmark Still Matters

Kimi K3 is not merely a large Chinese model attracting attention because of its parameter count. Independent testing supports a more substantial conclusion: it is competitive with frontier systems in automation, long-horizon knowledge work, and agentic execution.

That does not make it better than GPT-5.6 Sol or Claude Fable 5 overall. It also does not erase the measured increase in hallucination. What changed on July 28 is deployment: the promised weights, license, model card, and technical report are now public.
K3 is therefore one of the most capable open-weight options in its class, but the phrase needs two footnotes. Its license is custom rather than MIT, and its 1.56 TB repository plus 64+ accelerator recommendation put self-hosting beyond an ordinary development team.
My verdict: use the API to find out whether K3 actually improves your tasks. Move to the weights only when control, customization, or data boundaries justify becoming the infrastructure provider yourself.