AI tools·Chat models·China
Kimi
Started with huge context; K2/K3 moved into open flagship models and long coding. Kimi Code and Kimi Work complete the suite.
Compiled through 2026-09
- Vendor
- Moonshot AI
- Started
- 2023
- China availability
- Works in ChinaDirectly reachable from mainland China
- Pricing tier
- Free tier
- Current flagship
- Kimi k1.5 / K3
- Website
- www.kimi.com
Evaluation, Pros & Cons
Editorial score based on public benchmarks and community feedback — not a third-party benchmark result
Key Advantages
- Long-context pioneer with seamless 2M-token document analysis
- k1.5 multimodal reasoning + Explorer mode provide premier deep-research capabilities
Limitations & Caveats
- Multi-file software engineering is less focused than dedicated coding agents
Overview
Kimi first stood out for book-length context. 2025’s K2 became a global open-weight contender; K2.6 / K2.7 Code pushed agents and programming. July 2026’s K3 (~2.8T MoE, KDA linear attention, 1M context, native vision) is slated for open weights. Surfaces: kimi.com, mobile apps, Kimi Work, and Kimi Code in the terminal.
Direction
Moonshot is bundling long context, open flagship weights, and coding agents. After K3, inference partners and the open ecosystem are the bottleneck; Code / Work turn the model into a daily workbench.
Version history
Jul 2026
LatestKimi K3
2.8T MoE, 1M context, open-weight plan.
View details
New capabilities
- ≈2.8T MoE with KDA linear attention
- 1M tokens plus native vision
- Open-weight plan
What got stronger
- Long context moved from a product trick to an architecture choice
Version notes
K3 is the July 2026 flagship story: huge MoE, million-token context, promised open weights. Treat shipping and quotas as official-only. Product surfaces remain kimi.com plus Code / Work.
May 2026
Kimi K2.6 / K2.7 Code
Long-horizon agents and coding SKUs.
View details
New capabilities
- K2.6 long-horizon agents
- K2.7 Code as a coding SKU
What got stronger
- Coding stopped being a side effect of the chat model
Version notes
K2.6 / K2.7 split agents and coding into named SKUs and launched Kimi Code / Work. Moonshot was walking into the IDE, not only a reader.
Jul 2025
Kimi K2 open weights
Open flagship; global weight tier.
View details
New capabilities
- K2 open flagship weights
- Entered the global open-weight first tier
What got stronger
- From a nice app to a self-hostable base
Version notes
K2 open weights were Moonshot’s turn: the product was no longer only kimi.com. Overseas builders started seeing Kimi on boards and in local deploys.
Mar 2024
Long-context upgrades
Multi-million-character reading.
View details
New capabilities
- Marketed multi-million-character reading
- Better long-doc summary, compare, and extract
What got stronger
- Research, contracts, and filings became more complete workflows
Version notes
This upgrade made “long” daily, not a demo: upload a stack, ask for diffs and extracts. Hallucinations and citation quality still bite, but the window itself was a moat.
Oct 2023
Kimi launch
Huge-context reading as the hook.
View details
New capabilities
- Book-length / long-PDF reading
- Web product, Chinese-first UX
What got stronger
- Could finish materials most models could not even load
Version notes
Kimi launched on “it can finish the document,” not exam scores. Early intelligence was ordinary; the product memory was sharp, and it set up K2/K3’s long-context DNA.