- 20
- July
Kimi is a chatbot and family of large language models (LLMs) from Moonshot AI, a Chinese AI startup based in Beijing. Its flagship models — Kimi K2 and Kimi K2 Thinking — are open-weight models whose weights can be downloaded and self-hosted, and they are counted among the most capable open-weight models of 2026. This article explains what Kimi is, how good it really is when you separate vendor-reported numbers from independent testing, and — most important for a Thai business — where your data goes when you use it.
In one line: Kimi is an open-weight AI model from China's Moonshot AI that is genuinely strong — "the most capable model from a China-based developer" at launch — but still trails leading US models (per NIST's independent testing). For a Thai business, the more useful question than "which model is smartest" is "where does our data go, and how easily can we switch models?"
Who Are Kimi and Moonshot AI?
Moonshot AI (月之暗面, "the dark side of the moon" — a nod to Pink Floyd's The Dark Side of the Moon) was founded in March 2023 by Yang Zhilin, a former Tsinghua University researcher who had worked at Meta and Google, together with co-founders Zhou Xinyu and Wu Yuxin. It is headquartered in Beijing. The Kimi chatbot first launched in October 2023, and Moonshot quickly became one of China's "AI Tiger" startups.
Its anchor investor is Alibaba, which led a roughly US$1B round in February 2024 (reported at around US$800M for roughly a 36% stake), followed by Tencent, Gaorong Capital and IDG Capital in later rounds. By mid-2026, Moonshot was reported to be raising around US$2B at a valuation reported anywhere from US$20B to US$31.5B (some talk of up to US$50B) ahead of a Hong Kong IPO targeted for H2 2026 or early 2027. Annual recurring revenue (ARR) was reported at around US$200M in April 2026 and later cited at around US$300M. These figures conflict and are not finalised — read them as "reported," not confirmed fact.
The Kimi Model Lineage: From Chatbot to K2 Thinking
What makes Kimi notable is not just the chatbot but the speed of its releases and its decision to publish model weights openly. The table below summarises the key milestones. This article anchors on K2 and K2 Thinking because they are the best-documented and independently tested.
| Version / Date | Key traits | Notes |
|---|---|---|
| Kimi (chatbot) Oct 2023 | 128K-token context (~200,000 Chinese characters), later raised | The first selling point: reading very long documents in one pass |
| Kimi K1.5 20 Jan 2025 | A reasoning model | Moonshot claimed parity with OpenAI o1 on math and coding (a vendor claim) |
| Kimi K2 Jul 2025 | 1T total params / 32B active, Mixture-of-Experts; context expanded to 256K (K2-Instruct-0905, 9 Sep 2025) | Open-weight under a modified MIT license — one of the largest open-weight models |
| Kimi K2 Thinking 6 Nov 2025 | Same 1T/32B MoE base, but a step-by-step reasoning model; executes 200–300 sequential tool calls; exposes its reasoning trace via API; ships INT4 quantisation | The version this article cites for benchmark numbers |
| K2.5 / K2.6 / K3 2026 | K2.5 (Jan 2026) adds vision; K2.6 focuses on coding; K3 (Jul 2026) reportedly ~2.8T params | Iteration continues rapidly, but newer benchmark claims come mostly from marketing sources — treat with caution |
Kimi is optimised for long context, agentic tool use, coding/software engineering, and — for K2 Thinking — step-by-step reasoning. That profile means it is frequently compared with another well-known Chinese model, DeepSeek.
How Good Is It — Separate "Vendor Numbers" From "Independent Testing"
This is the heart of the article. When reading AI model news, the most common mistake is to treat numbers the vendor reported itself as if they were neutral facts — when in reality they were measured by the model's owner, in conditions the owner chose. Below we clearly separate two sets. The first is Moonshot's own reported numbers for Kimi K2 Thinking (from the official kimi.com blog).
| Benchmark (vendor-reported by Moonshot) | Kimi K2 Thinking score |
|---|---|
| SWE-Bench Verified (real software bug-fixing) | 71.3% |
| LiveCodeBench v6 (coding) | 83.1% |
| Humanity's Last Exam (with tools) | 44.9% |
| GPQA-Diamond (PhD-level questions) | 84.5% |
| AIME 2025 (competition math, with Python) | 99.1% |
| BrowseComp (web browsing) | 60.2% |
These look very impressive, but do not take them as neutral, central values — they were measured and defined by the developer. The right question is: "when an independent body re-measures, what does it get?" The answer comes from CAISI (Center for AI Standards and Innovation) under NIST, a US government body, which published its evaluation on 12 December 2025.
| Independent test (NIST/CAISI) | Kimi K2 Thinking | GPT-5 | Claude Opus 4 |
|---|---|---|---|
| SWE-Bench Verified (Moonshot's own claim: 71.3%) | 56.2 | 63.0 | 66.7 |
| CVE-Bench (cyber) | 50.5 | 65.6 | 66.7 |
| Math SMT 2025 | 93.1 | 91.8 | — |
| OTIS-AIME 2025 | 84.3 | 91.9 | — |
NIST's verdict is blunt: at release, Kimi K2 Thinking was "the most capable AI model from a PRC-based developer," but it "still lags behind leading U.S. models," and in cyber and math it was only "a modest improvement on DeepSeek V3.1." The point to highlight is the 71.3% (vendor claim) vs 56.2% (NIST measurement) gap on the same task — a concrete example of why a vendor benchmark is not the same as performance on your real workload. As for a comparison with Qwen, there is currently no clean, verified head-to-head, so we will not fabricate one (for the broader ranking picture, see our latest AI model comparison 2026).
Beware these two figures: (1) It is widely cited (e.g. by CNBC) that Kimi K2 Thinking was trained for only ~US$4.6M, but Moonshot's own researchers said this is "not an official number," so it must not be presented as fact. (2) The "DeepSeek moment" framing around Kimi is real, but many analysts argue its public impact was much more muted than DeepSeek R1's. In general, whenever you see a pretty vendor benchmark, always ask: "who measured it, and under what conditions?"
Another dimension independent testing catches — and which rarely appears in marketing benchmarks — is censorship. NIST found Kimi aligned with Chinese government positions about 26% of the time when asked in Chinese (rated "highly censored"), dropping to around 7% in English. This may not directly affect accounting or inventory work, but it matters a great deal if you plan to use the model to summarise news, analyse policy, or handle geopolitically sensitive content.
Open Weights: Kimi's Real Strength
What makes Kimi genuinely interesting for organisations is not its benchmark score but the fact that it is open-weight. Moonshot publishes the weights (moonshotai/Kimi-K2-Instruct and Kimi-K2-Thinking) on Hugging Face and the code on GitHub under a modified MIT license. That means you can download it and run it on your own servers — a major point for data governance (the same principle we cover in self-hosting Chinese models and the security of running models yourself).
The open-weight upside: When you run the weights in your own environment (on-premise, or a cloud you control), your prompt data never has to leave your organisation — it is not sent to servers abroad, and it is not used to train the provider's models. This is the fundamental difference between "using the vendor-hosted API" and "running the weights yourself," and it is why open-weight models are a serious option for organisations that care about open source vs commercial models.
That said, running a 1-trillion-parameter model yourself is neither easy nor cheap — it demands high-end hardware and skilled staff. Most organisations therefore use it via a hosted API, which brings us back to the data question. Reported K2 (0711) API pricing (from a price aggregator, which moves fast) is roughly US$0.55 per 1M input tokens and US$2.20 per 1M output tokens; newer versions cost more. Treat these as dated, illustrative figures only.
What Thai Businesses Must Consider: Where Does the Data Go?
When you use Moonshot's hosted API, the prompts and data you send are processed on China-based servers. According to the service's privacy policy, data may be transferred or stored on servers outside your country without naming a specific country, and user content may be used to train or optimise the provider's models (this is attributed to the policy). For Thai organisations subject to the Personal Data Protection Act (PDPA), sending customers' or employees' personal data abroad for processing requires a lawful basis and the ability to account for how that data is used (which connects to data subjects' right of access).
Data-governance cautions: (1) Data residency — using the hosted API means prompts are processed on servers in China. (2) Training on your data — the policy states user content may be used to improve the models. (3) China's National Intelligence Law (Article 7) can compel Chinese companies to provide data access — stated here neutrally as the letter of the law; how much of a "risk" that is remains analysts' interpretation, not settled fact. This principle applies to any China-hosted model, not only Kimi (read more on China AI risks).
To be clear, this article does not say Kimi is "unsafe" or "off-limits." Kimi is a genuinely good open-weight option, and if you self-host the weights in your own environment, the residency and training concerns largely disappear. The point is that the right question for a business is not "which model is smartest" but "where does our data go, and how easily can we move to another model?" The table below summarises the questions to ask before deciding.
| Question a Thai business should ask | Why it matters |
|---|---|
| Is this personal data or a trade secret? | If so, cross-border processing needs a PDPA basis and self-hosting deserves serious consideration |
| Are we using the hosted API, or self-hosting the weights? | Decides whether data leaves the organisation — the single biggest governance dividing line |
| If we ever have to switch models, how easily can we move? | Measures vendor lock-in — well-structured data can always plug into a new model |
| Have we tested it on our own real workload yet? | Vendor benchmarks ≠ real performance — test against your own data and tasks |
Where Does an ERP System Fit In (Honestly)?
AI models change leaders every few months. Today it might be Kimi; tomorrow it could be Claude, Gemini, or a brand-new Chinese model (see the pace of competition in the AI model war 2026). Betting on "which model will win" is therefore a volatile bet. What is not volatile — and pays off no matter which model wins — is your own clean, structured operational data. This is where a back-office system like ERP fits in: not because ERP is AI, but because ERP is the layer that keeps your data "clean, structured, and yours."
Saeree ERP stores core operational data — accounting, inventory, partners and budgets — in a single PostgreSQL database with an audit trail. That dataset is a portable asset that belongs to you, not tied to any one AI vendor. When the time comes to build on AI, you will have data ready to plug into whichever model fits best at that moment — a self-hosted Kimi, a commercial model, or something else. That hedges vendor lock-in and data-residency risk at the same time.
Being honest about scope: Saeree ERP does not claim to be an AI tool, and it does not embed Kimi or any model. Its AI assistant is still in development (training) — not yet a usable feature. What ERP can do today, and what matters most, is to get your data ready so you can choose any AI model from a position of leverage, without rushing to lock yourself into any single provider.
"The question that matters for a business is not 'which AI model is smartest' — that answer changes every few months. The real question is 'where does our data go, and how easily can we switch models?' Models change; your clean data stays with you."
- The Saeree ERP Team
References
- NIST / CAISI — CAISI Evaluation of Kimi K2 Thinking (independent testing, 12 Dec 2025)
- Moonshot AI — official Kimi K2 Thinking blog (vendor-reported benchmark numbers)
- Hugging Face — moonshotai/Kimi-K2-Thinking (open weights + license)
- Wikipedia — Moonshot AI (company history, founders, funding)
- Wikipedia — Kimi (chatbot) (chatbot and model evolution)
- Interconnects (Nathan Lambert) — 5 Thoughts on Kimi K2 Thinking (independent analysis)
Last verified 20 July 2026.
Interested in an ERP system for your organisation?
No matter which AI model leads, clean data unified in one place is the foundation that always pays off. Talk to the Grand Linux Solution team about consolidating your operational data for the AI era — free, no obligation.
Request a Free DemoTel 02-347-7730 | sale@grandlinux.com
