Verdict: Kimi K3 — Moonshot AI's 2.8-trillion-parameter open-weight model with a 1-million-token context window — is the first free, downloadable AI that handles entire document libraries, long coaching transcripts, and multi-step coding in a single conversation without losing the thread. For builders and small teams, the real unlock isn't the benchmark scores; it's the combination of massive context, native vision, and always-on reasoning that lets you replace a patched-together stack of "paste-it-in-chunks" tools with one model that takes the whole pile at once. The catch: you still need the API or serious GPU hardware to run it, and independent testing has flagged a hallucination rate worth respecting.
Last verified: 2026-07-31 TL;DR:
- Kimi K3 is a 2.8T-parameter Mixture-of-Experts model; only 16 of 896 experts activate per token, so it feels fast despite its size.
- 1-million-token context — feed it whole codebases, full call transcripts, or months of notes in one go.
- Open weights under the modified MIT "Kimi K3 License" on Hugging Face; commercial use permitted.
- Benchmarks: 93.5% GPQA Diamond, 91.2% BrowseComp, 88.3% Terminal-Bench 2.1 — competitive with paid frontier models.
- API pricing: $3 / million input tokens, $15 / million output tokens ($0.30 cache-hit rate).
- Volatile: pricing, limits, and benchmark numbers change fast — re-check before committing.
- Pricing/limits change often — last checked 2026-07-31.
What is Kimi K3 and why does the parameter count actually matter?
Kimi K3 is Moonshot AI's flagship model, released July 16, 2026 (API) with full open weights landing July 27, 2026 on Hugging Face under the moonshotai/Kimi-K3 repository. It is a 2.8-trillion-parameter Mixture-of-Experts (MoE) transformer — the largest open-weight model ever released, roughly 75% larger than DeepSeek V4 Pro's 1.6T parameters. Moonshot calls it the world's first "open 3T-class" model.
The parameter count matters for a practical reason, not just a marketing one: more parameters mean the model stores more knowledge and patterns, so it can handle harder reasoning, broader knowledge, and longer-horizon tasks without losing coherence. But the MoE architecture means you don't pay the full cost of all 2.8T on every request — only 16 of 896 expert sub-networks activate per token, roughly 1.8% of the total, giving you "the power of a giant brain at the speed of a much smaller one." Moonshot reports this makes the model about 2.5× more efficient per token than its predecessor Kimi K2.
Sources: Moonshot AI Kimi K3 launch blog (kimi.com/blog/kimi-k3); Hugging Face model card (huggingface.co/moonshotai/Kimi-K3); VentureBeat coverage, July 16, 2026.
How does Kimi K3 compare to GPT-5.6 Sol and Claude Opus 4.8?
On independent testing by Artificial Analysis, Kimi K3 scores an Intelligence Index of roughly 57 — placing it fourth among all frontier models, behind Claude Fable 5 (~60) and GPT-5.6 Sol (~59), and narrowly ahead of Claude Opus 4.8 (~56). It is the first open-weight model to place this high.
Where it leads is coding and frontend work: #1 on LMArena's Frontend Code Arena (1,679 Elo, beating Claude Fable 5), with the strongest open-weight GPQA Diamond score ever published at launch (93.5%).
| Benchmark | Kimi K3 | What it measures | Source |
|---|---|---|---|
| GPQA Diamond | 93.5% | Graduate-level science reasoning | Moonshot blog / Vals AI |
| BrowseComp | 91.2% | Web browsing and answer-finding | Moonshot blog |
| Terminal-Bench 2.1 | 88.3% | Real coding tasks in a terminal | Moonshot blog / BenchLM |
| Frontend Code Arena | #1 (1,679 Elo) | Human-judged web dev quality | LMArena |
| GDPval-AA v2 | 1,668 | Real-world knowledge work (44 occupations) | Artificial Analysis |
Cost per completed task (as measured by Artificial Analysis): ~$0.94 for Kimi K3, vs ~$1.04 for GPT-5.6 Sol and ~$1.80 for Claude Opus 4.8 — lower because the model uses fewer tokens to reach an answer, despite a higher per-token rate than its predecessor.
For a deeper comparison of frontier models head-to-head, see our GPT-5.6 Sol vs Claude Opus 5 comparison. For Moonshot's funding context and what K3 means strategically, read Moonshot AI Raised $3.5 Billion at $35 Billion — What Kimi K3 Means for You.
How much does Kimi K3 cost to use?
Kimi K3 is priced at $3.00 per million input tokens and $15.00 per million output tokens, with a cache-hit input rate of $0.30 per million. This is roughly the Claude Sonnet tier — a notable increase over Kimi K2.6's $0.95/$4.00 pricing — but lower per completed task than GPT-5.6 Sol or Claude Opus 4.8 thanks to token efficiency.
If you self-host the open weights, the model itself is free under the Kimi K3 License (a modified MIT license that permits commercial use, self-hosting, and fine-tuning). But running 2.8T parameters requires serious hardware: a minimum of 8× H100 80GB GPUs for experimental loads, and 64+ accelerators for production inference. The download is ~594 GB in MXFP4 format.
Source: codersera.com pricing guide (July 17, 2026); AIToolsRecap self-hosting notes (July 27, 2026); Moonshot platform pricing page.
Can you actually run Kimi K3 locally?
For most people, no — not the full model. The 2.8T parameter footprint requires a datacenter-class GPU cluster. Consumer hardware (an RTX 4090, a Mac Studio, a single workstation) cannot load the model even in quantized form.
Your practical options are:
- The Kimi API at platform.kimi.ai — the fastest way to use K3 today. Call model ID
kimi-k3and pay per token. - Managed inference providers — Together AI, Fireworks AI, and similar platforms were expected to host K3 within 48 hours of the weight release. Useful if you want OpenAI-compatible endpoints without signing up with Moonshot directly.
- Self-host on your own GPUs — only if you have 8+ H100s and a serving plan. vLLM with Kimi Delta Attention prefill-cache support shipped alongside the weights.
For a step-by-step guide to running K3 inside a coding agent (Claude Code, Cursor, etc.), see our companion piece: How to Run Kimi K3 in Claude Code: 3 Routes, Real Costs, and Which One Wins. And for the full architecture deep-dive on K3's internals, read Kimi K3 Explained: How Moonshot's 2.8T Open-Weight Model Changes the AI Compute Math.
What makes Kimi K3 different from other open models?
Three architectural innovations distinguish K3 from both closed frontier models and earlier open-weight releases:
Kimi Delta Attention (KDA)
Older models slow down as you feed them more text — the attention mechanism has to scan an ever-growing window, like searching one page in a disorganized stack. KDA is a hybrid linear attention mechanism that keeps decoding fast even at million-token scale. In plain terms, the model stays quick when you hand it a huge pile of documents rather than getting bogged down halfway through. Moonshot published this as open research (arXiv:2510.26692) before shipping it in K3.
Attention Residuals (AttnRes)
A drop-in replacement for standard residual connections that Moonshot says delivers consistent scaling gains — the model learns better as it gets larger, without the diminishing returns that plague some scaling efforts. Also published as open research (arXiv:2603.15031) and available on Moonshot's GitHub.
Stable LatentMoE
The MoE design activates only 16 of 896 expert networks per token. This is what lets a 2.8T model run at the cost and speed of a roughly 50B-parameter equivalent. Moonshot's framing: you call in only the specialists you need for each job instead of waking the whole office on every question.
Sources: Moonshot AI arXiv papers (2510.26692, 2603.15031); Hugging Face model overview; rejoicehub.com technical guide.
How to use Kimi K3 for real work: 4 prompt workflows
The benchmarks tell you K3 is capable. The real question for anyone running a business or building a product is: what can you actually do with it that you couldn't do before? The 1-million-token context window changes the economics of certain workflows because you stop spending time chunking, pasting, and re-summarizing — and start handing the model the entire context at once.
Here are four practical workflows, each with a ready-to-use prompt. These are distilled from hands-on use and serve as templates — adapt the specifics to your own business.
Workflow 1: Landing page copy that explains your offer clearly
The problem: most landing pages read like feature lists because no one has the patience to explain the actual value in plain language. A model with enough context can hold your entire offer, pricing, customer questions, and competitive positioning in mind at once.
Prompt template:
Design a new landing page for [your product or service] that converts
visitors into customers. Show the real value of [your category] for
[your audience]. Explain the [key features: e.g., weekly coaching
calls, daily tutorials, roadmaps, community map]. Focus the whole
page on helping [your audience] get more [desired outcome — leads,
customers, traffic] through [your mechanism]. Keep the language
simple enough for a complete beginner to follow.
Why K3 specifically: its 1M context lets you paste your entire existing site, customer support transcripts, and competitor pages into the same conversation so the model writes copy informed by your full context, not just the short prompt. Its native vision also means you can hand it a screenshot of your current page and it understands the layout, not just the text on it.
Result: a full page draft ready to edit, structured around outcomes rather than features.
Workflow 2: Turn long recordings into ready-to-post content
The problem: turning a 90-minute coaching call or customer interview into social content means either writing each post yourself or pasting chunks into a tool that forgets what happened three minutes ago. With a 1M context, K3 can hold the entire transcript in one pass.
Prompt template:
Take this [type: coaching call / customer interview / podcast]
transcript and turn it into [N] short posts for social media.
Each post should explain one clear tip from the transcript in
simple language. Focus only on tips that help a [your audience]
[save time / bring in more customers / solve a problem] using
[your topic]. Keep each post under 100 words and easy for a
beginner to follow.
Why K3 specifically: because it holds the whole transcript at once, it catches details smaller-context models miss halfway through. You stop losing the middle of long conversations.
Result: five (or however many you asked for) ready-made posts pulled from a real recording — nothing lost in the middle.
Workflow 3: A support guide built from your own community's questions
The problem: new customers ask the same questions over and over. Writing a help guide takes time you don't have, and generic AI tools produce generic answers. K3 can read your entire backlog of community discussions and FAQs in one shot and synthesize a guide that sounds like it came from your team.
Prompt template:
Read through these frequently asked questions and past community
discussions from [your community or product]. Build a simple support
guide that answers the most common member questions about getting
started, using [your key features], and [common sticking points].
Write it in plain, friendly language a brand new member could
understand right away.
Why K3 specifically: paste every Discord thread, support ticket, and FAQ doc you have — K3 reads them all in one conversation and cross-references answers across sources. A smaller model would need the inputs pre-filtered and chunked.
Result: a ready-made help guide built from real member questions, written in your community's voice, that a newcomer can read start to finish in minutes.
Workflow 4: Research synthesis across large document sets
The problem: when you're evaluating a market, a competitor, or a technology, the useful material doesn't fit in a single prompt — it's a pile of earnings reports, blog posts, product docs, and PDFs. Most tools make you summarize each one separately and then re-summarize the summaries.
Prompt template:
Read the following [N] documents: [paste or attach full text of
reports, articles, product docs, financials]. Synthesize the key
findings into a structured brief covering: (1) what [the market /
technology / competitor] actually is, (2) the strongest evidence for
and against [your hypothesis], (3) the three most important things
a [your role: builder / operator / investor] should pay attention
to in the next 90 days, and (4) any claims you cannot verify from
these sources alone — flag them clearly.
Why K3 specifically: the 1M context window holds roughly a small library. You feed it the whole pile at once and it cites back across sources. The instruction to flag unverifiable claims is important for trust — K3's hallucination rate has been independently tested at around 51% on certain adversarial benchmarks, a rate not disclosed in Moonshot's own charts, so explicit grounding instructions matter.
Source on hallucination rate: AIToolsRecap, July 27, 2026, citing independent testing.
What are the limitations and risks of using Kimi K3?
K3 is genuinely impressive, but honest evaluation means naming the caveats:
- Hallucination rate: independent testing (AIToolsRecap, July 27, 2026) found a 51% hallucination rate on certain adversarial tests — a figure not included in Moonshot's benchmark charts. Always verify load-bearing facts before you act on or publish K3's output.
- Hardware cost for self-hosting: the open weights are free, but running them requires 8× H100 80GB minimum and 64+ accelerators for production. For most teams the API is the practical path; self-hosting only makes sense if you have the GPUs or a hard data-residency requirement.
- Training-data provenance: the White House has publicly accused Moonshot of using fraudulent accounts to harvest Claude training data (reported July 2026, unresolved). The license covers your right to use the weights, not questions about how the training data was obtained. Factor this into vendor risk assessments for enterprise deployment.
- Always-on reasoning mode: K3 ships with reasoning always on ("thinking mode"), which means every request runs the full chain-of-thought. This is good for quality but means you can't turn it off to save tokens on simple tasks.
- License specifics: the "Kimi K3 License" is MIT-like but not plain MIT. It includes commercial gates — for example, hosting K3 as a paid Model-as-a-Service above a $20M annual revenue threshold requires a separate deal with Moonshot. Read the actual license file before building a commercial product on top of K3.
Sources: AIToolsRecap (July 27, 2026); kimi-k2.org license analysis; Hugging Face model card.
What this means for you
If you run a small business or agency: Kimi K3's 1-million-token context is the feature that actually changes your workflow, not the parameter count. Try it for any task where you currently spend time chunking, summarizing, and re-summarizing — long transcripts, big document sets, whole-support histories. Start with the API at platform.kimi.ai; you pay per token and can stop any time. Use the four prompt templates above, adapted to your offer and audience.
If you're a developer or builder: K3 is the first open-weight model that competes with the top proprietary systems on capability, not just price. If you self-host, budget for real infrastructure. If you build on the API, watch the hallucination rate and add a verification layer. The modified MIT license is permissive for most commercial uses, but read the file before shipping a product — especially if you plan to resell model access.
If you're evaluating open-source AI for your stack: the open-weight landscape moved fast in 2026, and K3 is the current ceiling for what's freely downloadable. But the gap between "strongest open model" and "strongest model" still exists — Claude Fable 5 and GPT-5.6 Sol remain ahead on broad intelligence. K3's advantage is that you can own it, fine-tune it, and run it on your terms.
FAQ
Q: Is Kimi K3 really free and open source?
A: The model weights are freely downloadable under the Kimi K3 License (a modified MIT license) from Hugging Face at huggingface.co/moonshotai/Kimi-K3. Commercial use is permitted, but there are gates — for example, offering K3 as a paid Model-as-a-Service above $20M annual revenue requires a separate agreement with Moonshot. The hosted API at platform.kimi.ai is paid ($3 / million input tokens, $15 / million output tokens).
Q: How big is Kimi K3? A: Kimi K3 has 2.8 trillion total parameters — the largest open-weight model ever released. It uses a Mixture-of-Experts architecture with 896 experts, of which only 16 activate per token (roughly 1.8% of the total), so it runs at the efficiency of a much smaller model despite its size.
Q: What is Kimi K3's context window? A: 1 million tokens. That is roughly enough to hold a small library of documents, a full codebase, or months of notes in a single conversation without losing earlier context.
Q: How does Kimi K3 compare to GPT-5.6 or Claude? A: On the Artificial Analysis Intelligence Index, K3 scores ~57 — fourth among all frontier models, behind Claude Fable 5 and GPT-5.6 Sol, and narrowly ahead of Claude Opus 4.8. It leads on the Frontend Code Arena (human-judged web dev quality) and posts the strongest open-weight GPQA Diamond score (93.5%) at launch.
Q: Can I run Kimi K3 on my own hardware? A: Only if you have serious GPU infrastructure. The minimum experimental load is 8× H100 80GB GPUs; production requires 64+ accelerators. The download is ~594 GB. For most users the API is the practical path. Consumer hardware (RTX 4090, Mac Studio) cannot run the full model.
Q: What is Kimi Delta Attention? A: KDA is a hybrid linear attention mechanism developed by Moonshot AI that keeps decoding fast even at million-token context lengths. It was published as open research (arXiv:2510.26692) before being shipped in K3. In practice, it means the model does not slow down as you feed it more text — a problem with older attention designs.
Q: Who makes Kimi K3? A: Moonshot AI (月之暗面), a Beijing-based AI company founded in March 2023 by Yang Zhilin and colleagues. The company is backed by Alibaba, Tencent, HongShan, IDG Capital, and others, and was valued at $20 billion in its May 2026 Series D round.
Discussion
0 comments