NewChat Models

Kimi

Kimi K3.1 Pro High Speed

Best for Coding

Faster-serving K3.1 Pro variant with the same provider-described intelligence, 1M context and separate API pricing.

1M contextHigh speed

At a glance

Know the model before you prompt.

Kimi API specifications
Context window
1,048,576 tokens
Maximum output
1,048,576 request ceiling; shared with context
Inputs → output
Text + Images + Video → Text
Knowledge cutoff
Not verified

API model ID: kimi-k3.1-pro-highspeed

Capabilities & boundaries

What it supports. Where the limits are.

Tool support requires the appropriate API integration; a supported tool is not automatically active in every chat.

Supported API features and tools

  • A 1,048,576-token context for organizing repository material, source documents and visual evidence
  • Always-on thinking with low, high and max reasoning effort; max is the documented default
  • Streaming with separate reasoning_content and final-answer content deltas
  • Strict JSON Schema for final content, partial-prefix continuation, custom tool calls and dynamic tool loading
  • Automatic prefix caching with five-minute or one-hour TTL options and separate cache-write charges
  • A faster-serving variant: Kimi describes the same intelligence as Pro and advertises about 2–3× the speed

Before you choose

  • The shared quickstart specifies a 131,072 default max_completion_tokens budget and a 1,048,576 request ceiling. Input plus completion must fit the context; the ceiling does not guarantee a million-token final answer or exclude reasoning from the budget.
  • Visual understanding returns text. Blender, CAD, spreadsheet editing and web research require external tools and an integration that executes them; this model alone is not an image, video or 3D-file generator.
  • Public image URLs are not supported by the documented vision workflow. Use supported base64 content or uploaded ms:// file references in a content array. App upload support is not established by this API capability.
  • Kimi’s quickstart says its web-search tooling is being updated and is not recommended for near-term production workflows. This is provider-tool guidance, not an assessment of EZ Ai Assist’s separate search feature.
  • The 2–3× speed claim is provider-reported, not a guaranteed time to first token or end-to-end latency. Long inputs, reasoning, queueing and external tools affect actual wait time. No separate High Speed benchmark measurements are supplied here.
  • No verified knowledge-cutoff date is supplied. The October 24 documentation announces full weights for November 4; this guide does not claim they are already released or establish their eventual license.

Choose the reasoning effort

lowhighmax · default

Thinking cannot be disabled. Set the top-level reasoning_effort to low, high or max. Start with a bounded task and compare correctness, latency and billed usage across effort levels; lower effort reduces the reasoning budget rather than creating a non-thinking mode. The same-family High Speed variant changes serving speed, not the documented effort names.

  • Use the exact model ID kimi-k3.1-pro-highspeed with the Kimi Chat Completions API. These are official provider IDs, not independently verified EZ Ai Assist upstream mappings.
  • Use max_completion_tokens for the completion budget. The current API reference says it automatically lowers the generation limit when input plus the requested budget exceeds the context window. Output usage includes reasoning and answer tokens. Preserve the complete returned assistant message, including reasoning and tool history, in later turns.
  • The documented sampling values are fixed: temperature 1.0, top_p 0.95, n 1 and both penalties 0. Omit unsupported adjustments rather than copying another provider’s generic request settings.
  • With strict JSON Schema, parse final message.content, not reasoning_content. Validate both schema and factual values. For tool calls, execute permitted actions separately and return a result for each matching tool_call_id.
  • Dynamic tools require a complete definition in the documented system-message form and must remain in subsequent history. Formula tools also require a client-side execution loop; declaring a tool does not execute it.
  • Automatic caching applies to eligible repeated prefixes. Five minutes is the default TTL; select 1h through prompt_cache_options.ttl when appropriate. A partial block is a cache miss, and a cache hit is not another cache-write billing event.
  • Direct Kimi API access requires a successful top-up, with tiers controlling rate limits. That provider-account requirement is not an EZ Ai Assist subscription or purchase instruction.

Put it to work

Start with a more useful prompt.

Original examples from EZ Ai Assist. Adapt these to your task and the features available in your workspace.

Workflow 01

Triage one failing test

Keep the feedback loop focused and reviewable.

Use the failing test output and relevant code below to identify the smallest plausible cause. Separate observed facts from hypotheses. Propose one minimal change and one regression test, citing the exact supplied lines. Do not refactor unrelated code or claim to run commands. If evidence is missing, ask for the single most useful next observation.

Workflow 02

Review a visual iteration

Turn a screenshot into a short acceptance checklist.

Compare the supplied before-and-after screenshots with this design requirement. Identify only observable differences in spacing, hierarchy, text wrapping and alignment. Prioritize the three changes with the largest usability impact and give a mobile and desktop acceptance check for each. Do not infer click behavior or hidden implementation from a screenshot.

Workflow 03

Refine an evidence brief

Improve a draft without inventing new facts.

Revise this brief using only the numbered source excerpts below. Preserve supported claims, flag contradictions, and remove conclusions without evidence. Return a concise revised brief followed by a change log with source numbers. Keep unresolved questions explicit and do not claim to browse or verify material outside the supplied sources.

Developer reference

Kimi API pricing

These are Kimi API reference prices, not EZ Ai Assist subscription prices or actual customer charges.

View EZ Ai Assist plans
Kimi K3.1 Pro High Speed direct API reference · verified October 24, 2026 · USD per 1,000,000 tokens
Token typePrice
Input · cache miss$4.00
Input · cache hit$0.20
Output$20.00
Cache write · 5-minute TTL$4.00
Cache write · 1-hour TTL$8.00
  • Verified against the rendered official pricing table on October 24, 2026. Rates are USD per million tokens, exclude applicable taxes and can change. Input and output are separate billing items; context length does not create a different price tier in the documented K3.1 Pro pricing.
  • Cache writes have separate five-minute and one-hour rates. Cached input receives the cache-hit rate without an additional write charge on that hit. Do not sum every row as though every token incurs every charge.
  • High Speed has twice Pro’s listed uncached-input, output and cache-write rates, while both list $0.20 per million cached-input tokens. A speed claim alone does not establish better cost per successful task.

Common questions

A few things worth knowing.

How does Pro differ from High Speed?

Kimi describes High Speed as the same intelligence with about 2–3× the speed. It uses a distinct API ID and different reference rates. Treat that as provider positioning, not a guarantee of equal answers or fixed response times in every integration.

Can I turn thinking off?

No. The shared quickstart documents always-on thinking, with low, high and max reasoning effort and max as the default. Effort must be configured by the integration; mentioning a setting in a prompt does not prove an API field was sent.

Does 1M context mean a million-token answer?

No. The documented completion ceiling is a request budget shared with context, not a guaranteed final-answer length. Input, reasoning and other request constraints affect what remains for a response; the default completion budget is 131,072.

Can it produce a finished 3D scene or spreadsheet?

Its documented output is text. Kimi’s examples use tools and execution environments to produce artifacts. A chat response can propose code or instructions, but actual files require an integration that executes and validates the workflow.

Are the charts independent tests of EZ Ai Assist?

No. They are Kimi-published K3.1 Pro benchmark figures, not tests run by EZ Ai Assist and not separate High Speed evaluations. Tool setups and reasoning effort affect results; benchmark cost per task is not a customer charge.

Are model weights already available?

As reviewed October 24, 2026, the quickstart announces full model weights for November 4. An announced release is not confirmation of current availability. Check the official release and license before planning self-hosting.

Which features are available in my EZ Ai Assist workspace?

This guide is a provider reference. It does not independently verify current app routing, exposed reasoning controls, upload support or plan eligibility. Check your workspace and the EZ Ai Assist pricing page; direct Kimi API charges are a separate product.

Check the source

Official documentation

Specifications and API prices checked on . Example prompts and workflow advice are editorial guidance from EZ Ai Assist.

Same provider