Try Kimi K3
1M context • always-on reasoning • text + image input • tool calling
Kimi K3 Assistant
long-horizon coding • visual reasoning • knowledge work • structured output
Fast Response
Get instant answers powered by our optimized infrastructure
Privacy First
Your conversations are secure and never used for training
Premium Features
Sign in to unlock API access and unlimited conversations
Kimi K3 at a glance
These figures come from Moonshot AI's Kimi K3 launch materials. Kimrel is an independent access service and is not affiliated with Moonshot AI.
Total Parameters
2.8T
A sparse Mixture-of-Experts model in the three-trillion-parameter class
Context Window
1M
Up to one million tokens for large repositories, document sets, and long sessions
Active Experts
16 / 896
Kimi K3 routes each token through sixteen experts in its Stable LatentMoE design
Kimrel Minimum
3 Credits
Minimum charge for each successful request before token-based billing is applied
Why developers are evaluating Kimi K3
Kimi K3 combines a larger sparse architecture with long context, native vision, persistent reasoning, and agent-oriented API features for multi-step work.
Long-horizon software engineering
Kimi K3 is built for sustained engineering sessions. It can inspect broad repository context, follow detailed constraints, use terminal-style tools, and continue after an initial approach fails. Gate edits with tests and human review.
One-million-token context
Kimi K3 expands the context window to one million tokens for large codebases, specifications, research collections, and durable agent histories. Relevant context selection still matters for latency, cost, and quality.
Native visual understanding
Kimi K3 can reason over text and images in the same request. Screenshots, diagrams, charts, and interface states can become evidence inside a coding or analysis workflow. Kimrel accepts image input for this route but intentionally rejects video input.
Always-on deep reasoning
Kimi K3 always reasons before producing a final answer. The official API uses the top-level `reasoning_effort` field and currently documents `max` as the supported level. It does not use the K2.x `thinking` request object.
Tool calling and structured output
Kimi K3 supports function tools, required tool choice, dynamic tool definitions, JSON Schema output, and partial continuation. These controls suit typed workflows that need predictable arguments and validated results.
New sparse architecture
Kimi K3 combines Kimi Delta Attention, Attention Residuals, Gated MLA, and Stable LatentMoE. Moonshot AI reports roughly 2.5 times the overall scaling efficiency of Kimi K2, with sixteen of 896 experts activated for each token.
What the published Kimi K3 evidence shows
Moonshot AI's launch report mixes benchmark results with long-running case studies. The examples below are useful signals, not guarantees. Evaluate Kimi K3 with your own prompts, tools, repositories, latency targets, and correctness checks before production adoption.
DeepSWE — 67.3
Moonshot AI reports that Kimi K3 reaches 67.3 on DeepSWE with the mini-SWE-agent harness. The evaluation targets software-engineering agents rather than short code completion, and harness details still affect outcomes.
BrowseComp — 90.4
With the full one-million-token context and no context-management layer, Kimi K3 is reported at 90.4 on BrowseComp. Production research still depends on source selection, tool reliability, and verification.
MiniTriton compiler project
In a published case study, Kimi K3 developed a compact Triton-like system with a tile-level IR, optimization passes, PTX generation, and a working nanoGPT training pipeline—connected engineering across multiple layers.
GPU kernel optimization
Kimi K3 received up to 24 hours to profile, rewrite, and benchmark four GPU-kernel tasks. Moonshot AI says an early version also handled much of the team's late-stage kernel optimization work.
Forty-eight-hour chip design
A Kimi K3 proof of concept used open-source EDA tools to design and verify an accelerator in 48 hours. The simulation closed at 100 MHz and exceeded 8,700 decode tokens per second.
Research-to-code workflow
For an astrophysics case study, Kimi K3 reviewed twenty-plus papers, evaluated over 300 equations of state, generated 3,000-plus lines of Python, and built an interactive dashboard linking research, implementation, and presentation.
Where Kimi K3 fits best
Choose Kimi K3 when the work benefits from deep reasoning, broad context, visual evidence, and repeated tool use. Simpler prompts may be cheaper and faster on another route, so model selection should follow the task rather than the version number.
Repository-scale feature delivery
Give Kimi K3 the architecture, acceptance criteria, coding rules, and relevant source. It can trace work across routes, services, database code, and tests. Keep the agent in a reviewable branch and verify its claims.
Visual frontend debugging
Send Kimi K3 a screenshot together with component code, browser output, and the desired state. Native vision lets it connect visible spacing, hierarchy, overflow, and interaction problems to implementation details instead of guessing from text alone.
Large-document knowledge work
Kimi K3 can work across lengthy reports, contracts, specifications, and research collections within one large context. Ask for evidence-linked conclusions, conflicting claims, and explicit uncertainty so a long answer remains auditable rather than merely fluent.
Research with executable analysis
Use Kimi K3 when research must end in code, calculations, charts, or a dashboard. Separate source gathering, implementation, validation, and presentation so each stage can be checked independently.
Controlled tool-using agents
Kimi K3 can choose functions, inspect results, and continue across multiple turns. Production agents should expose narrow tools, validate every argument, cap iterations, record tool outcomes, and require approval before destructive or externally visible actions.
Technical design exploration
Kimi K3 can compare architectures, reason through constraints, and turn a design into an implementation sequence, especially when code, diagrams, logs, and written requirements must be considered together.
Call Kimi K3 through Kimrel
Kimrel exposes Kimi K3 through the same independent API-key flow as its other supported routes. Existing clients can keep their authentication and endpoint structure, then select the new model ID and account for its reasoning and multimodal behavior.
Use the exact model ID
Set `model` to `kimi-k3`. Kimi K3 uses the shared `KIMI_API_KEY` upstream configuration, so Kimrel does not introduce a second provider credential or a separate authentication path for this model.
Use reasoning_effort, not thinking
On `/v1/chat/completions`, Kimi K3 accepts `reasoning_effort: "max"`. Reasoning is always enabled, and the K2.x `thinking` object should not be sent. Reserve enough completion tokens for both reasoning tokens and the final response.
Choose either compatible endpoint
Use Kimi K3 through `/v1/chat/completions` or `/v1/messages`. The Anthropic-compatible route maps supported message and tool structures upstream, while K3 uses max reasoning by default.
Send text and supported images
Kimi K3 accepts text plus PNG, JPEG/JPG, WEBP, or GIF images on Kimrel. Although Moonshot AI documents native video understanding for the underlying model, this service deliberately rejects video files and video URL parts.
Let Kimrel normalize remote images
For Kimi K3, send an HTTP(S) image URL or a `data:image/...;base64,...` value. Kimrel fetches remote images, validates the type, blocks unsafe hosts, and converts them to base64. Original files are limited to 6MB and encoded payloads to 8MB.
Plan token-based credit usage
Kimi K3 costs 300 input and 1,500 output credits per million tokens, with a three-credit successful-request minimum. Kimrel uses reported usage when available and its documented fallback otherwise.
Straight answers about Kimi K3
These answers separate Moonshot AI's published model capabilities from the specific boundaries of the independent Kimrel service.
What is Kimi K3?
Kimi K3 is Moonshot AI's 2.8-trillion-parameter flagship model for long-horizon coding, knowledge work, visual reasoning, and tool-using agents. It combines Kimi Delta Attention, Attention Residuals, and a sparse Mixture-of-Experts design with a one-million-token context window.
Is Kimi K3 available through Kimrel now?
Yes. Kimi K3 is registered as `kimi-k3` in Kimrel's chat, playground, model-list, billing, and API routing layers. It uses the same Kimrel API key as the other supported models and can be called from both compatible endpoints.
Does Kimi K3 support images and video?
The underlying Kimi K3 model has native visual capabilities, and Moonshot AI documents image and video understanding. Kimrel currently accepts text and images only. Video content is intentionally rejected so the public service behavior remains aligned with its validated forwarding pipeline.
How should Kimi K3 reasoning be configured?
Kimi K3 always reasons. On the OpenAI-compatible endpoint, use the top-level `reasoning_effort` field with `max`, which is also the current default. Do not send the K2.x `thinking` object, and preserve the complete assistant message in multi-turn tool workflows.
Can Kimi K3 call tools and return JSON?
Yes. Kimi K3 supports function calling, `tool_choice`, dynamically loaded tools, JSON mode, strict JSON Schema output, and partial continuation in the upstream API. Applications must still validate model-produced arguments and enforce their own authorization boundaries.
Is Kimrel the developer of Kimi K3?
No. Kimi K3 is developed by Moonshot AI. Kimrel is an independent platform that provides a compatible access layer, credit billing, remote-image normalization, and its own service controls. Kimrel is not affiliated with, endorsed by, or sponsored by Moonshot AI.
Put Kimi K3 on a demanding workflow
Try Kimi K3 in chat, then use the Kimrel API for long-context, image-aware, tool-assisted applications.