Moonshot
Chat with Kimi K3 online
Moonshot's open-weight flagship that sees: screenshots, charts and video frames in, a million tokens of context, reasoning always on.
1M tokens
Context
Not publicly disclosed
Knowledge cutoff
32,768 tokens
Max output
Fast
Speed
Straight from Kimi K3
Three prompts, three unedited answers, generated on September 2, 2026 through NinjaChat. Click the prompt to copy it and run it yourself.
Write a Python function that merges overlapping intervals, plus three short tests. Keep the whole answer under 40 lines of code.
def merge_intervals(intervals):
"""Merge overlapping intervals. Each interval is [start, end]."""
merged = []
for start, end in sorted(intervals):
if merged and start <= merged[-1][1]:
merged[-1][1] = max(merged[-1][1], end)
else:
merged.append([start, end])
return merged
def test_merge_intervals():
# Overlapping and disjoint intervals
assert merge_intervals([[1, 3], [2, 6], [8, 10], [15, 18]]) == [[1, 6], [8, 10], [15, 18]]
# Touching intervals merge; unsorted input handled
assert merge_intervals([[5, 7], [1, 4], [4, 5]]) == [[1, 7]]
# Empty and single-interval edge cases
assert merge_intervals([]) == [] and merge_intervals([[2, 3]]) == [[2, 3]]
test_merge_intervals()
print("All tests passed.")
The function sorts intervals by start time, then walks through once, extending the last merged interval when there's overlap. Runs in O(n log n) time.
Read: We ran the same 3 prompts on 12 AI models — Unedited answers, latency and price for every model, measured through the NinjaChat API in September 2026.
Served by Moonshot via OpenRouter as kimi-k3. Answers vary run to run; these are one take each, not the best of several.
What you’ll use it for
Code plus screenshots
Paste the error and the screenshot; K3 reads both and fixes the right thing.
Very long context
A million tokens: whole repos, log archives, document sets in one thread.
Agentic tasks
Multi-step tool use with reasoning that stays on across the whole job.
Visual reasoning
Charts, diagrams, UI mockups and video frames as first-class input.
Prompts to steal
Tuned to this model — click any line to copy.
About Kimi K3
Kimi K3 is Moonshot AI's flagship, released in July 2026 as the successor to Kimi K2.5. It is a 2.8 trillion parameter mixture-of-experts model that activates 16 of its 896 experts per token, built on Moonshot's Kimi Delta Attention and attention residuals, and it comes with native vision, a 1 million token context window and open weights. On NinjaChat it is included in every plan, with no Kimi membership needed.
Two things set K3 apart in daily use. It sees: images and video frames go straight into the conversation, so it can read a screenshot of a failing build, a chart from a report or a frame from a clip and reason about it alongside your text. And it keeps the whole thing in view: a million tokens of context means a large repository, a long thread of logs and tests, or a stack of documents fits in one session without summarizing on the way.
Moonshot built K3 for long-horizon work. Reasoning is always on, with an effort level you can raise for the hard cases, and the model is unusually good at coordinating tools over many steps: navigating a codebase, running commands, reading the output and adjusting. That makes it a strong pick for coding agents, front-end work that mixes code with visual judgment, and research tasks that stitch many sources together.
It is also open weights, which matters if you care about where your model comes from or want the option of running it yourself later. On NinjaChat it sits next to Claude, GPT-5.6 and Gemini, so the practical question is which one fits the task: K3 for visual-plus-code problems and enormous contexts, the others when their ecosystems or styles suit you better.
Full access to K3 in Moonshot's own app sits behind a paid membership; here it is part of the same NinjaChat plan as everything else.
How to use Kimi K3
From zero to your first result in under a minute.
01
Create a NinjaChat account and choose a plan
02
Open chat and pick Kimi K3 from the model list
03
Drop in screenshots, files or a whole folder of context
04
Ask for the full job, not a snippet; K3 plans across steps
05
Compare the same prompt on Claude or GPT-5.6 with one click
Tips for better results
- Give it the screenshot, not a description of it
- For hard problems, ask it to reason more before answering
- Use it for front-end work where code and visuals meet
- Keep long projects in one thread; the context is there for that
Kimi K3 vs the alternatives
Where it wins, and where it doesn't.
vs Kimi K2.5
ViewNative vision, a much larger expert pool and stronger long-horizon agentic work
K2.5 remains a capable, cheaper text model
vs Claude Opus 5
ViewOpen weights, native video-frame input and a comparable context window
Opus 5 leads on agentic coding benchmarks and tool polish
vs GPT-5.6 Sol
Open weights and strong multimodal reasoning at lower cost
Sol has the deepest reasoning tier in OpenAI's family
Questions, answered.
Beyond chat — images & video
The same subscription covers the image studio and video generation.
Kimi K3, and 50 more.
One subscription covers every model on NinjaChat, Kimi K3 included.
