← Back to Browse
MIXED
60/100
UNBOUGHT SCORE
CONFIDENCE 100%
Mixed reads
K

Kimi

AI-powered chatbot for long-context conversations and file analysis
MIXEDAI Build
Pricing
Freemium
Learning curve
Moderate
Time to value
Days

Kimi Family

62
Kimi Code
44
Kimi-K2.5
Gathering
data
Kimi Work

First-hand reads coming soon. We're collecting real user feedback.

Visit WebsiteSave to Stack
BEST FORSmall BusinessMid-MarketSolopreneur
COMMUNITY SENTIMENT
PositiveNegative
WORKS FOR

Research-heavy workflows, long document analysis, users who need 128K+ context windows.

DOESN'T WORK FOR

Simple chat use cases; less brand recognition than ChatGPT or Claude in Western markets.

Related Tools

Deep Research96.3% CONF
73/100
AI agent that autonomously conducts deep-dive web research on complex topics.
View
DeepSeek100% CONF
71/100
Open-source large language models from DeepSeek AI, optimized for reasoning and …
View
Retell AI95% CONF
69/100
Platform for building AI voice agents that answer and make phone calls, used by …
View
Manus59.3% CONF
69/100
Autonomous AI agent that plans and executes multi-step tasks
View

FIRST-HAND READS

What people reported, in their own words. Every line links to where it was said.

  • “Kimi K3 through the CLI if sol is trying to test something security related and gets policy stopped many times in a row”

    POSITIVE · features · Reddit · 2026-08-15

  • “Hermes on Kimi 3. Burning right through my weekly tokens on 2 projects.”

    NEGATIVE · price · Reddit · 2026-08-13

  • “Only Pi, OpenClaw and Kimi stamp a real format or protocol version that an external reader can dispatch on.”

    POSITIVE · features · Reddit · 2026-08-13

  • “Kimi K3 will happily do security work, and they do it pretty well”

    POSITIVE · performance · Hacker News · 2026-08-13

  • “Kimi ends up being quite expensive and rarely better than GPT Sol or Opus 5”

    NEGATIVE · price · Hacker News · 2026-08-07

  • “all the things I cannot discuss with our four frontier champion shill models, I can with GLM, or Kimi, though it takes work.”

    POSITIVE · Hacker News · 2026-07-29

  • “Kimi 3 is more expensive than Opus 5 or Sol on Cursor? Pretty sure Kimi 3 sucked up a good chunk of my ultimate plan in a few prompts.”

    NEGATIVE · price · Hacker News · 2026-07-29

  • “showed a massive uplift for Kimi K3 and this tracks a much improved 3D capability”

    POSITIVE · performance · Hacker News · 2026-07-17

  • “Kimi K3 just found the source of a bug that Fable 5 hasn't been able to pinpoint in multiple attempts”

    POSITIVE · performance · Hacker News · 2026-07-17

  • “it ran out of free credits. So I was not able to go on a Journey with Kimi K3”

    NEGATIVE · Hacker News · 2026-07-17

  • “neither gemini nor kimi were able to write something working”

    NEGATIVE · performance · Hacker News · 2026-07-10

  • “built a custom authentication layer with logging and limits for Dify MCP with just 2 prompts in Kimi”

    POSITIVE · ease of use · Hacker News · 2026-07-03

  • “I used a number of times Kimi K2.7 and I was disappointed. It would run in circles for things that Claude would do in one shot.”

    NEGATIVE · performance · Hacker News · 2026-07-02

  • “It costs $0.006 on Kimi K2.7, and you get to see the whole raw reasoning trace.”

    MIXED · price · Hacker News · 2026-07-02

  • “use something like Kimi, GLM or Deepseek for a fraction of the price of Anthropic et al”

    POSITIVE · price · Hacker News · 2026-06-30

  • “Kimi k2.7 code seems to have extra training against cyber security capabilities. Last one, k2.6 was a lot stronger at cyber but obviously the Kimi team improved over time, so this is not the best they can do but no one will be able to get the best anymore”

    NEGATIVE · features · Hacker News · 2026-06-29

  • “Kimi k2.6 with Kimi code beats opus models at finding vulnerabilities”

    MIXED · performance · Hacker News · 2026-06-28

  • “You can tell Kimi or whatever these days to write you a GTK4 mobile apps for a lot of pet use cases and it will.”

    POSITIVE · features · Hacker News · 2026-06-26

  • “kimi k2.6/7 inside pi finds more issues and inside kimi code finds the most”

    MIXED · performance · Hacker News · 2026-06-25

  • “Deepseek, Kimi, MiMo started giving me multiple options but advocating strongly that I should either accept this deficiency, or don't use the cookies version”

    NEGATIVE · performance · Hacker News · 2026-06-25

  • “kimi-2.5-a-bunch-of-specs is going to behave more consistently than whatever tweaks Claude is making to Sonnet this week”

    POSITIVE · reliability · Hacker News · 2026-06-24

  • “I tried kimi but it generally spins it's wheels extensively in it's thinking tokens.”

    NEGATIVE · performance · Hacker News · 2026-06-20

  • “KIMI K2 0905 completely hallucinated a real name for me (I don't work on Pygame!)”

    NEGATIVE · performance · Hacker News · 2026-06-19

  • “kimi is already there, it has started to not believe the user and do what it was taught just like claude models (which is helpful when you don't care about implementation specifics or performance/security)”

    NEGATIVE · Hacker News · 2026-06-18

  • “the latest versions of Kimi, MiMo, and GLM (pre-5.2) had already surpassed Sonnet in terms of output quality for a fraction of the price”

    POSITIVE · performance · Hacker News · 2026-06-17

Have You Used This Tool?

Sign in to share your experience with this tool.

Sign in