Fact-checked: August 5, 2026 · Model status: active flagship · Evidence level: official documentation and model card, plus a separately reported single-pass API pilot; no reliability benchmark or cross-model comparison.
Kimi K3 is Moonshot AI’s flagship open-weight model for long-horizon coding, reasoning and knowledge work. The hosted Kimi API model ID is kimi-k3, its maximum context window is 1,048,576 tokens, and it accepts text, image and video inputs. K3 always uses thinking, but API users can choose low, high or max reasoning effort.
K3 is the highest-capability option in the current Kimi line, but it is not the cheapest. Its official output rate is $15 per million tokens, compared with $4 for Kimi K2.7 Code and Kimi K2.6. Use our Kimi model comparison if you are still choosing.
Kimi K3 at a glance
| Field | Verified value |
|---|---|
| Current status | Active; Kimi’s flagship as of August 4, 2026 |
| Release date | July 16, 2026 |
| Global Kimi API model ID | kimi-k3 |
| Kimi Code model IDs | k3 and k3-256k |
| Official open-weight repository | moonshotai/Kimi-K3 |
| Architecture | Mixture of Experts using Kimi Delta Attention and Gated MLA |
| Total / activated parameters | 2.8T / 104B |
| Context window | 1,048,576 tokens on the Global API |
| Hosted API input | Text, image and video |
| Thinking | Always enabled |
| Reasoning effort | low, high, max; Global API default is max |
| Tool calling | Supported |
| Structured output | JSON mode and strict JSON Schema supported |
| License for weights | Kimi K3 License |
The specification table comes from Kimi’s API guide and official K3 model card. Product access and default settings can differ between the Global API and Kimi Code.
What is different about Kimi K3?
K3 expands both model scale and usable context. Moonshot reports 2.8 trillion total parameters, 104 billion activated parameters, 896 experts with 16 selected per token, and a hybrid attention stack combining Kimi Delta Attention (KDA) with Gated Multi-head Latent Attention. A MoonViT-V2 vision encoder provides native visual capability.
Those architecture figures explain how the model is built; they do not prove that it is better for every prompt. A smaller model can be faster, cheaper or more consistent on a constrained workflow. The decision should be based on task-level evaluation, not parameter count alone.
One-million-token context
The Global Kimi API lists a 1,048,576-token context window for kimi-k3. This is useful for very large repositories, long document collections and agent histories that exceed the 262,144-token windows of K2.7 Code and K2.6.
Four caveats matter:
- Capacity is not recall. A model accepting a long prompt does not guarantee that it will retrieve or reason over every detail equally well.
- Input and output share practical limits. Leave enough room for the response and tool messages instead of filling the window with input.
- Agent software may compact first. Kimi Code, Codex, Claude Code or another harness may summarize history according to its own context policy.
- Long prompts cost more. A one-million-token cache miss costs $3 in input tokens before any output or paid tools.
Kimi says context caching is automatic. A new request can attempt a prefix-cache hit only after an earlier request has more than 256 prompt tokens, and the long prefix must remain unchanged. A published cache-hit rate should therefore come from actual usage data, not an assumption.
Thinking mode and reasoning effort
K3 cannot run in a non-thinking mode. On the Global Kimi API, use the top-level reasoning_effort field:
| Setting | Practical starting point |
|---|---|
low | Routine questions, simple transformations and latency-sensitive steps |
high | Complex but bounded coding or analysis |
max | Hard reasoning, long-horizon agents or evaluation runs where quality takes priority |
These descriptions are workload guidance, not measured performance guarantees. Kimi’s Global API documentation says the default is max; the separate Kimi Code documentation lists high as its default. Record the product and effort level whenever you compare results.
For multi-turn conversations and tool calls, Kimi instructs developers to send the complete assistant message back in the next request, including reasoning_content and tool_calls. Keeping only the visible answer can break preserved-thinking behavior and make an agent comparison invalid.
Multimodal input: what is supported?
The hosted Kimi API guide includes text, image and video examples. It also documents several restrictions:
- Vision message
contentmust be an array of typed objects. - Public image URLs are not supported in the documented K3 flow; use base64 data or an uploaded
ms://<file-id>reference. - Video can be uploaded through the Files API and then referenced in a request.
- Media consumes tokens, so resolution, duration and selected video frames can affect cost.
The open-weight model card’s architecture table labels the modality as text and image while its descriptive section says K3 understands video. The hosted API clearly exposes video input; do not assume that every self-hosted runtime reproduces the official video pipeline without checking its implementation.
Kimi K3 API pricing
Official list prices checked August 4, 2026:
| Billing component | Price per 1M tokens |
|---|---|
| Input with cache hit | $0.30 |
| Input with cache miss | $3.00 |
| Output | $15.00 |
Prices exclude applicable taxes. The API is pay as you go and is separate from Kimi membership. Kimi’s K3 quickstart says the model is unlocked after a successful top-up, with a $1 minimum; cumulative top-ups also determine account rate-limit tiers.
Cost calculator
Use this formula for a first estimate:
cost = (cached input × 0.30 + uncached input × 3.00 + output × 15.00) ÷ 1,000,000
Example: 500,000 uncached input tokens plus 50,000 output tokens costs approximately $2.25 before tax: $1.50 input + $0.75 output. Retries, tool calls and additional turns add to the total.
See Kimi API pricing for product-level billing, rate-limit and caching guidance.
Where can you use Kimi K3?
Global Kimi API Platform
Use kimi-k3 with the OpenAI-compatible endpoint documented at https://api.moonshot.ai/v1. The platform exposes pay-as-you-go billing, API keys, a Playground and account-level rate limits. This is the model surface described by the pricing figures on this page.
Kimi Code
Kimi Code uses different IDs:
k3: K3 with access up to 1M context for eligible higher-tier members.k3-256k: a fixed 256K K3 option intended to reduce quota consumption; Kimi’s documentation says it does not support video input.
Kimi Code access depends on membership tier. Its quota usage is not interchangeable with Global API token billing, and its documented default reasoning effort is high, not the Global API’s max.
Open weights and self-hosting
Moonshot publishes K3 weights at Hugging Face and code/report material on GitHub. The model card recommends vLLM, SGLang and TokenSpeed deployment paths.
K3 is a very large model. “Open weight” does not mean it can be served economically on ordinary consumer hardware. Quantization, available accelerators, tensor parallelism, inference engine and chat template all influence feasibility and output quality. The Kimi K3 License—not an assumed open-source license—governs the released weights.
Features relevant to agents and applications
K3’s official API documentation lists:
- automatic context caching;
- custom tool calls and
tool_choice; - dynamic tool loading;
- JSON mode and strict JSON Schema output;
- Partial Mode for continuing from a supplied prefix;
- streaming with separate reasoning and final-answer deltas;
- multimodal messages;
- official tools through Kimi’s Formula workflow.
K3 does not access the web by default. Kimi’s current API documentation provides official web-search tools that applications can add, so search remains a separate tool dependency rather than inherent K3 knowledge. Verify current model/mode compatibility, tool-token use and billing before production use; the first-party pages currently disagree on the per-call search fee.
Vendor-reported benchmark snapshot
The following scores come from Moonshot’s K3 model card. They are not independent Kimi AI Guide results.
| Benchmark | Kimi K3 reported score | Relevant setup note |
|---|---|---|
| GPQA Diamond | 93.5 | K3 at max effort |
| DeepSWE | 67.5 | Kimi Code harness; model card also notes 67.3 under mini-SWE-agent on the official leaderboard |
| Terminal-Bench 2.1 | 88.3 | Kimi Code harness |
| BrowseComp | 91.2 | Context compaction at 300K; card reports 90.4 with full 1M and no context management |
| MMMU-Pro | 81.6 / 83.4 | Without / with Python-tool augmentation |
Moonshot reports K3 results with reasoning_effort="max" and temperature 1.0. Some competitor scores in the same table use different agent harnesses, cited leaderboards or fallback behavior. Several listed benchmarks are in-house. These differences make the table evidence of the vendor’s evaluation, not a universal ranking.
Best-fit use cases
K3 is a sensible candidate when:
- a repository or evidence set genuinely exceeds 256K tokens;
- the task combines visual evidence with code or tool use;
- an agent must sustain a long sequence of engineering or knowledge-work steps;
- structured output and a large tool inventory are required;
- quality matters enough to justify K3’s higher output price.
K2.7 Code may be a better first choice for cost-sensitive software engineering. K2.6 may be better for general-purpose workloads that need optional non-thinking responses.
Important limitations
- K3 always thinks; there is no true instant/non-thinking switch.
- The full 1M context is a maximum capacity, not a guarantee of perfect long-context accuracy.
- API output is substantially more expensive than K2.7 Code or K2.6.
- Preserved reasoning history increases implementation complexity and context usage.
- Video support is clearest on the official hosted API; third-party and self-hosted behavior may differ.
- Web search is an optional external tool path, not an inherent guarantee of current or cited knowledge; availability, mode compatibility and billing require separate verification.
- Benchmark scores depend on effort, tools, harness, context management and evaluator design.
- Like any generative model, K3 can produce incorrect code, unsupported citations and confident factual errors.
Our independent testing status
On August 5, 2026, we completed a single-pass Global Kimi API pilot under frozen protocol v1.1. All 15 calls completed and reported kimi-k3: eight tool-call cases passed 8/8, the code repair passed all 5/5 fixed tests without an out-of-scope file change, the vision response matched 6/6 fields, and the five context requests matched 20/20 target fields from approximately 32K through a request with 950,149 prompt tokens.
The usage-derived token cost was $5.687397 at the dated official rates. A read-only balance check 13 minutes 41 seconds after completion showed a $5.68739 deduction, a $0.000007 difference. This is a balance observation, not a tax invoice. Every case and context band ran once on one account, date and environment; the result is not a reliability estimate, a universal K3 score or evidence that every near-limit prompt will be recalled correctly. See the complete Kimi K3 real-world pilot and our testing methodology.
Frequently asked questions
Is Kimi K3 free?
The Global API is not a permanent free API tier: Kimi’s documentation requires a successful top-up of at least $1 to unlock K3. Kimi consumer or membership access follows separate quotas and can change independently.
What is the Kimi K3 API model ID?
Use kimi-k3 on the Global Kimi API. In Kimi Code, the documented IDs are k3 and k3-256k; they are not interchangeable with the Global API ID.
Can Kimi K3 process video?
Yes on the official hosted API, which documents video-file input. The k3-256k Kimi Code option does not support video, and self-hosted runtimes should be verified separately.
Can Kimi K3 turn thinking off?
No. K3 always reasons. The Global API lets you reduce effort with reasoning_effort="low", but that is not the same as disabling thinking.
Is Kimi K3 open source?
Kimi describes K3 as open weight and publishes the model files. The weights use the specific Kimi K3 License. Review that license for your use case instead of assuming Modified MIT terms from earlier Kimi models.