Understand images and video
Kimi K3 natively understands text, images, and video. Dreammmer's first live version will keep the workflow focused on one image plus a question and return a text answer.
Official release verified · July 16, 2026
Kimi K3 Online: Multimodal Reasoning and Coding
Kimi K3 is Moonshot AI's 2.8T-parameter flagship for long-horizon coding, visual understanding, knowledge work, and deep reasoning. Explore its 1M context, Max, High, and Low modes, API status, and planned Dreammmer experience.
Dreammmer is an independent product, not the official Kimi website. Kimi K3 has been released, but our provider is not connected yet. The preview below sends no text or image. Kimi K3 understands visual input and returns text; it is not presented here as an image generator.
Kimi K3 multimodal workspace
Planned runtime: kimi-k3 · Max reasoning
Reasoning effort
Max
Hardest tasks
High
Balanced depth
Low
Faster answers
Optional image
JPG or PNG · visual understanding, not image generation
Inputs will be enabled only after the provider model ID, vision format, limits, and output behavior are verified.
Planned tasks
Turn image into promptReview UI screenshotReview codePlan knowledge work
When access opens, prompts, uploaded images, and answers will follow Dreammmer's safety controls. Verify important claims and code before use.
Kimi response
After a verified provider is connected, the real answer, model ID, and reasoning effort will appear here. This preview does not simulate a response.
What it is built for
What users can do with Kimi K3
The page centers on the tasks that match Kimi K3's published design, with a clear boundary between model capability and Dreammmer's initial experience.
Long-horizon coding
The model is designed for extended coding and agent tasks. The first page will support focused code review before any repository or terminal workflow is considered.
Work across 1M tokens
The official API lists a 1M-token context window. The initial interface will not imply that large-file upload is available until the complete request path is tested.
Complete knowledge work
Use Kimi K3 to analyze supplied material, reason through constraints, and plan structured work. Tool calls require the surrounding application to provide and control the tools.
Kimi models and modes
Kimi K3 Max, High, Low, K2.7 Code, and K2.6 compared
K3 uses one official API model ID with three reasoning-effort settings. Earlier Kimi models remain useful for more specialized or lower-latency work.
Search nameOfficial configurationBest fit
Kimi K3 Max
kimi-k3 · reasoning_effort=max
The most difficult coding, visual, reasoning, and agent tasks where depth matters more than latency.
Kimi K3 High
kimi-k3 · reasoning_effort=high
Strong reasoning with a more practical balance of response time and depth.
Kimi K3 Low
kimi-k3 · reasoning_effort=low
Faster responses for simpler analysis, extraction, and everyday coding questions.
Kimi K2.7 Code
kimi-k2.7-code · 256K context
High-speed coding and agent workflows when K3's full reasoning depth is unnecessary.
Kimi K2.6
kimi-k2.6 · 256K context
General chat and multimodal understanding on the previous model generation.
All Kimi K3 search variants stay on this canonical page. Every future live answer will show the actual model ID and reasoning setting used.
Confirmed model facts
What changed in Kimi K3
Official claims are identified as such. Independent reporting is linked separately so release facts and outside evidence are not mixed together.
Last verified: July 20, 2026
Release and weights
Moonshot AI released Kimi K3 on July 16, 2026 across Kimi.com, Kimi Work, Kimi Code, and its API. The company announced that full weights would follow by July 27; Dreammmer will update this after verification.
Architecture and scale
Official materials list 2.8T total parameters, KDA plus Attention Residuals, and Stable LatentMoE activating 16 of 896 experts. Moonshot reports about 2.5× overall scaling efficiency versus K2.
API and developer features
The OpenAI-compatible API uses model ID kimi-k3 and documents streaming reasoning, JSON Schema output, partial mode, custom tools, dynamic tool loading, and automatic context caching.
Current limits
K3 always uses thinking, supports low, high, or max effort, and fixes several sampling parameters. Vision input does not accept public image URLs, and the official guide does not recommend web search for production while it is being updated.
A clear two-model workflow
Understand an image with Kimi. Create it with Seedream.
Kimi K3 and Seedream solve different jobs. Keeping those jobs separate makes the page more useful and the model claim more accurate.
Generate images with Seedream 5.0 ProUse Kimi to analyze a reference or improve a prompt, then use Dreammmer's live Seedream workflow to create the image.Open image generatorCompare all model experiencesSee which Dreammmer pages are live, which are waiting for provider access, and which task each model is designed to handle.Open model directory
Kimi K3 questions
What to know before using Kimi K3
Availability, visual capability, model choices, reasoning modes, and open-weight status in one place.
Can I use Kimi K3 on Dreammmer now?
Not yet. Kimi K3 is officially available, but Dreammmer has not connected and verified a provider. Text and image inputs are intentionally disabled, and the preview sends no data.
Can Kimi K3 generate images?
Kimi K3 is documented as understanding text, images, and video and returning text. This page does not claim that it generates images. Use Dreammmer's Seedream 5.0 Pro workflow when the desired output is an image.
What is the difference between Kimi K3 and Kimi K2.7 Code?
Kimi K3 is the stronger general flagship with a 1M-token context and native multimodal reasoning. Kimi K2.7 Code has a 256K context and is positioned for efficient coding and agent workflows.
What do Max, High, and Low mean in Kimi K3?
They are reasoning_effort settings for the same kimi-k3 model. K3 always thinks: Max prioritizes depth, High balances depth and latency, and Low is the faster option for simpler work.
Is Kimi K3 open source or available for local use?
Moonshot AI announced that full Kimi K3 weights would be released by July 27, 2026. Because that date had not arrived at our last verification, this page does not yet claim that the weights are downloadable.
Sources and verification
Official Kimi sources support release, architecture, API, mode, and pricing details. Associated Press reporting supplies a separate outside view of the launch.