Kimi K3 (2026): 2.8T Model, Pricing, API & Weights

Moonshot AI released Kimi K3 on July 16, 2026: a 2.8-trillion-parameter model with a one-million-token context window, native vision capability, and an API available now. Moonshot calls it the first open 3T-class model; full weights are promised by July 27, 2026.

This guide separates Moonshot's launch claims from independent measurements, covers official API pricing and access, and explains what the pending-weights date means. Last source check: July 17, 2026. Read the primary sources: Moonshot's Kimi K3 launch post, Kimi API documentation, and Artificial Analysis' live model profile.

What Kimi K3 is

Kimi K3 is Moonshot AI's July 2026 flagship, a 2.8 trillion parameter sparse mixture-of-experts model with a one million token context window. Moonshot calls the architecture "Stable LatentMoE," and it is aggressively sparse: only 16 of 896 experts fire on any given token, which puts activation under 2 percent of the total parameter count.

Because only a small subset of experts runs for each token, a sparse MoE does not perform inference as though all 2.8T parameters were active at once. Moonshot has not published an active-parameter total or the technical report yet, so any precise active-parameter estimate should be treated as unconfirmed.

Moonshot attributes K3's scale to Kimi Delta Attention and Attention Residuals, plus Stable LatentMoE. It reports roughly 2.5× higher overall scaling efficiency than Kimi K2; that is a vendor claim, not an independently reproduced result. The technical report is still pending, so the safest reading is that the architecture is promising but not yet fully documented in public.

K3 supports text and image input and text output in its first-party API. At launch, Moonshot says K3 uses maximum thinking effort by default, with additional effort settings planned later. That makes it sensible to measure both quality and token usage on your own prompts before choosing it for a production workflow.

For builders, the operational detail matters more than launch nicknames: Moonshot's documented API model ID is kimi-k3. Avoid relying on unconfirmed variant names or hardware requirements until the model card, weights, and technical report are published.

Kimi K3 benchmarks: where it wins and loses

Kimi K3 is strong enough to merit serious testing, but one leaderboard cannot establish an overall winner. The launch table below is Moonshot's self-reported set, measured at temperature 1.0 with maximum reasoning. Harnesses differ between models, so treat it as directional vendor evidence rather than a clean apples-to-apples independent comparison.

The table shows a mixed frontier profile rather than a universal win. K3 is competitive on long-horizon coding and browser-style work in Moonshot's evaluation, while other models lead on several tasks. That is why a model choice should start with the workload you actually run—not an average of unrelated leaderboards.

Independent measurements are the more useful second opinion. On July 17, Artificial Analysis scored K3 at 57 on its Intelligence Index and reported a 1M context window, $3 input and $15 output pricing, and 62 output tokens per second. Its live rankings change as new results arrive. K3 also led the Frontend Code Arena when checked on July 17, but that leaderboard measures a specific kind of blind frontend preference—not all software engineering.

There are real tradeoffs. Artificial Analysis currently characterizes K3 as slower than its comparable-model average and unusually verbose: 130M output tokens across its Intelligence Index versus a 63M peer median. Moonshot also documents three caveats: K3 is sensitive to preserved thinking history, can be overly proactive when instructions are ambiguous, and still has a noticeable user-experience gap versus the leading proprietary models. Test reliability, latency, and cost—not benchmark score alone.

Kimi K3 API pricing and how to access it

The official Kimi K3 API costs $3.00 per million cache-miss input tokens, $0.30 per million cached input tokens, and $15.00 per million output tokens, with a 1M-token context window. These are Moonshot's published rates as checked July 17, 2026. Recheck them before launch: model prices, cache behavior, and rate limits can change quickly.

The per-token price is not the whole cost. K3 currently uses maximum thinking effort by default, and independent testing found it unusually verbose. For a production estimate, run representative prompts, log cache-hit rate, input and output tokens, latency, and failure retries. That gives you a cost-per-successful-task number instead of a misleading list price.

Where you can use it today:

• Kimi.com and Kimi apps. Moonshot lists K3 in Kimi's web and mobile experiences.
• Kimi Work. Moonshot's desktop knowledge-work product supports K3.
• Kimi Code. Update the CLI, then select Kimi K3 with the /model command.
• Official API. Use the kimi-k3 model on the Kimi API Platform.

What you cannot do yet is download the full weights. Moonshot says it will release them by July 27, 2026. Until the download and licence are actually public, describe K3 accurately as a hosted model with announced open weights—not a model you can self-host today.

Moonshot recommends deploying K3 on supernode configurations with 64 or more accelerators. That puts it far beyond practical local hardware for individuals. Even after weights arrive, most teams will use a managed provider or dedicated infrastructure rather than a desktop machine.

Using Kimi K3 with the OpenAI SDK

Kimi documents an OpenAI-compatible Chat Completions API, so migration is straightforward for many applications. Set your SDK base URL to https://api.moonshot.ai/v1, keep your Moonshot key on the server, and call kimi-k3. Review Kimi-specific parameters before assuming feature parity.

What to test before adopting Kimi K3

K3 is worth testing for long-horizon coding, agentic workflows, and visual tasks—but build the evaluation around your real work. The model is new, its public evidence mixes vendor and third-party results, and its weights are still pending.

• Quality: use a fixed set of real tickets, documents, and frontend tasks with acceptance criteria.
• Reliability: test ambiguous prompts and multi-turn sessions; Moonshot warns that K3 can be sensitive to thinking history and overly proactive.
• Cost: track reasoning, output, cache-hit, and retry tokens for each successful task.
• Latency: include tool calls and end-to-end completion time, not just tokens per second.
• Deployment: wait for the actual weights and licence before making an open-weights or self-hosting decision.

Kimi K3 vs GPT-5.6 vs Claude Fable 5

Kimi K3 is a serious option for frontier coding and agentic work, but its strongest claim is specific: it led a frontend-focused arena when checked on July 17. That does not settle every coding, reasoning, reliability, or cost question. Compare models on the same tasks, configuration, and budget.

The honest verdict: Kimi K3 has credible early evidence on coding and agentic work, plus a 1M context window and an accessible API. Its announced open-weights status is not a completed self-hosting option until the weights and licence are public. If you want the OpenAI side of this comparison, read our GPT-5.6 Sol, Terra, and Luna guide.

Pick by task, not brand: test Kimi K3 for long-context, frontend, and agentic work; test competing models on the failure modes that matter to your product; and use measured cost per accepted result to decide.

Model rankings move quickly. A platform like Casagbic lets you describe the app while its agents use frontier models behind the scenes, so you can focus on product outcomes instead of rewriting your workflow whenever a leaderboard changes.

Frequently asked questions

The bottom line on Kimi K3

Kimi K3 is one of the most consequential model launches of 2026: 2.8 trillion parameters, a one-million-token context window, a live API, and an announced weights release by July 27. It also comes with real operational tradeoffs—verbosity, benchmark volatility, and a deployment footprint beyond consumer hardware. Treat it as a model to test today, not a claim to accept on launch-day headlines.

The durable advice is the same as always: validate a model on the work you actually run before you standardize on it, cache reusable context to control cost, and re-check the leaderboards often, because they keep moving.

Want to skip the model-picking entirely and just get an app built? Try it on Casagbic →