Verdict: Kimi K3 from Moonshot AI is the largest open-weight AI model ever released — 2.8 trillion parameters, a 1-million-token context window, and native vision — and its weights are free to download under the Kimi K3 License. But for 99% of people, the practical path is the free web chat or the $3/$15 per-million-token API, not self-hosting. The model ranks #1 on the Frontend Code Arena (1,679 Elo) and #3 on the Artificial Analysis main board, trailing Claude Fable 5 and GPT-5.6 Sol in overall benchmarks. The real story is not just size — it is architecture: a new Mixture-of-Experts design that activates only 104B of its 2.8T parameters per token, making a frontier-scale model practical to serve.
Last verified: August 1, 2026 — Pricing, benchmarks, and hardware requirements confirmed against primary sources. Pricing and model versions change often; re-check before committing.
- What it is: 2.8T-parameter MoE model from Moonshot AI, released July 16, 2026 (API), weights published July 27, 2026
- Context window: 1,000,000 tokens (1M) — enough for an entire codebase or book in one prompt
- API price: $3.00/M input tokens, $15.00/M output tokens, $0.30/M cached input
- Free access: Yes — kimi.com web chat, switch from K2.6 to K3 in the model dropdown
- Self-hosting: Technically free to download, but requires 64+ accelerators (datacenter class)
- Strongest area: Frontend coding (#1 on Frontend Code Arena, beating Claude Fable 5)
- Weakest area: Overall UX and hard SWE benchmarks still trail Claude Fable 5
What Is Kimi K3 and Why Does the Open Source Release Matter?
Kimi K3 is an open-weight AI model built by Moonshot AI, a Beijing-based lab. It launched on July 16, 2026 as an API and web chat model, and the full model weights were published on GitHub and Hugging Face on July 27, 2026 under the Kimi K3 License. It is the first open model to reach the 3-trillion-parameter class, making it the largest open-weight model ever released as of August 2026. The open weights mean anyone can download, deploy, and fine-tune the model — subject to the license terms — without paying per-token fees to a closed API provider.
The significance goes beyond the parameter count. Until Kimi K3, the strongest AI models — OpenAI's GPT-5.6, Anthropic's Claude Fable 5, Google's Gemini — were all closed systems. You could only use them through the vendor's API or web app. Kimi K3 is the first model with open weights that genuinely competes with closed frontier models on specific benchmarks, particularly frontend coding. That changes the buy-versus-host calculation for teams that need data sovereignty, cost control at scale, or the ability to customize a frontier-class model.
How Does Kimi K3's Architecture Work?
Kimi K3 uses a Mixture-of-Experts (MoE) architecture. Think of it as a team of 896 specialist models ("experts") inside one system. For every token the model generates, only 16 of those 896 experts are activated, plus 2 shared experts. This is why the total parameter count is 2.8 trillion but only about 104 billion parameters are "doing work" at any given moment. That gap is what makes a model this large practical to serve — you get the capacity of a 2.8T model with the inference cost of a ~100B model.
Two architectural innovations distinguish K3 from its predecessor, Kimi K2:
Kimi Delta Attention (KDA): A new attention mechanism that changes how information flows across long inputs. 69 of the model's 93 layers use KDA. This is what enables the 1-million-token context window without the memory explosion that plagues standard attention at that length.
Stable LatentMoE with Attention Residuals (AttnRes): The MoE routing framework that selects which experts to activate. Moonshot reports a roughly 2.5x scaling efficiency improvement over Kimi K2, meaning the model gets more capability per unit of compute than a straightforward scale-up would deliver.
The model also has native vision — it processes text and images within the same model, not through a bolted-on adapter. The vision encoder is MoonViT-V2 (401M parameters). Training used MXFP4 quantization for weights and MXFP8 for activations, meaning the model was trained with quantization awareness baked in, not applied post-hoc.
How Fast and Smart Is Kimi K3 Compared to Other Models?
On Moonshot's own published benchmarks (using reasoning_effort="max", temperature=1.0), K3 hits strong numbers but does not universally beat the closed frontier. Here is how it compares on key benchmarks, based on the official GitHub evaluation table:
| Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol | What it measures |
|---|---|---|---|---|
| GPQA Diamond | 93.5 | 92.6 | 94.1 | Graduate-level science reasoning |
| Terminal-Bench 2.1 | 88.3 | 88.0 | 88.8 | Terminal/CLI coding tasks |
| BrowseComp | 91.2 | 88.0 | 90.4 | Web browsing + agentic research |
| SWE-Marathon | 42.0 | 35.0 | 39.0 | Long-horizon software engineering |
| ProgramBench | 77.8 | 76.8 | 77.6 | General programming |
| DeepSWE | 67.5 | 70.0 | 73.0 | Deep software engineering |
| FrontierSWE | 81.2 | 86.6 | 71.3 | Frontier-level SWE tasks |
| MCPMark-Verified | 94.5 | 87.4 | 92.9 | MCP tool-use agentic tasks |
The pattern: K3 wins on long-context, agentic, and terminal-based tasks. Claude Fable 5 wins on the hardest software engineering benchmarks (FrontierSWE, DeepSWE) and overall conversational polish. Moonshot itself acknowledges this in their blog, stating that K3 "still trails Claude Fable 5 and GPT-5.6 Sol" overall.
On the Frontend Code Arena — a leaderboard run by Arena.ai that ranks models by blind human preference on real frontend coding tasks — K3 debuted at #1 with 1,679 Elo, ahead of Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618. This was a 17-place jump from K2.6, which sat at #18. The result was based on 483,895 blind human votes and K3 won 6 of 7 frontend domains. This is the strongest signal that K3 excels at frontend and visual coding work.
How Can You Actually Use Kimi K3 Right Now?
There are four ways to access Kimi K3, each for a different type of user:
1. Free web chat (simplest path)
Go to kimi.com, sign in with a free account, and switch the model from K2.6 (the default) to K3 in the dropdown. No API key, no payment, no setup. This is how most people should start.
Important: on a free account, avoid the "K3 swarm" mode — it burns through your free token allowance far too quickly to finish a real task. Select K3 directly and give it a specific, concrete task.
2. API (for developers)
The Kimi API is OpenAI-compatible, meaning you can point existing tools at it with minimal config changes. The model ID is kimi-k3.
Pricing (verified against Moonshot's official pricing page at platform.kimi.ai):
| Usage | Rate per 1M tokens |
|---|---|
| Input (cache miss) | $3.00 |
| Input (cache hit) | $0.30 |
| Output | $15.00 |
| Web search add-on | $0.004 per invocation |
For comparison, the previous flagship Kimi K2.6 costs $0.95/$4.00 per million tokens with a 256K context window. K3 costs more but delivers a 4x larger context window and stronger benchmarks. If you're cost-sensitive and don't need 1M context or K3's specific strengths, K2.6 is still available.
3. Desktop and terminal tools
Moonshot offers Kimi Work (desktop app) and Kimi Code (terminal-based). In terminal tools, you select the model with a /model command. These are useful if you want K3 integrated into your development workflow without going through a browser.
4. Self-host the weights (for teams with serious hardware)
The full model weights are on GitHub (MoonshotAI/Kimi-K3) and Hugging Face under the Kimi K3 License. The model can be served with vLLM, SGLang, or TokenSpeed. But the hardware requirements are extreme — see the next section.
What Does It Actually Cost to Self-Host Kimi K3?
Self-hosting is free to license but expensive to run. Here is the honest breakdown:
The model weights in MXFP4 format require roughly 1.4 TB of memory just to load. Moonshot's own recommendation is to serve it on setups with 64 or more accelerators. For context, a single NVIDIA H100 GPU has 80 GB of memory. You would need approximately 18 H100s just to hold the weights — and that is before accounting for KV cache, batching, and inference overhead. Practical deployments need dozens of GPUs with fast interconnect (NVLink or InfiniBand).
| Self-hosting path | Hardware needed | Best for |
|---|---|---|
| Buy your own cluster | 64+ GPUs (H100/H200/B200), NVLink fabric, ~1.4 TB+ VRAM | Regulated industries that must keep data in-house |
| Rent GPUs by the hour | Equivalent cluster on a cloud provider | Bursty workloads, occasional self-hosting needs |
| Mac Studio (unified memory) | Not viable for the full model — even 512 GB unified memory is far short | Use a smaller open model instead |
For teams evaluating whether the API or self-hosting makes economic sense: the API at $3/$15 per million tokens is far cheaper up to very high volumes. Self-hosting only wins financially when your token usage is so high that per-token API costs exceed the cost of running a datacenter GPU cluster — and the crossover point is enormous given the hardware footprint.
The non-cost reason to self-host is data sovereignty: Moonshot is a Beijing-based company, so API data travels to Chinese infrastructure. For regulated US firms under HIPAA, client contract restrictions, or data-residency rules, self-hosting the open weights is the compliance answer.
What Are the Key Gotchas When Using Kimi K3?
Two operational details that trip people up on day one:
Gotcha 1: Always start a fresh session — never switch models mid-conversation
Kimi K3 was trained in "preserved thinking history" mode. The full reasoning chain from each turn must be passed back to the model on the next turn. If you switch an existing chat from K2.6 to K3 halfway through, the thinking history is missing or incompatible and output quality drops sharply. You will think the model is broken. It is not — you just violated its training assumptions. Always start a new session with K3.
Gotcha 2: K3 is proactive — set clear boundaries or it builds what it thinks you meant
K3 was trained on long, complex tasks. When your instructions are vague, it will make decisions for you instead of asking for clarification. If you need it to operate within strict constraints, spell those constraints out explicitly in your system prompt or agent configuration file. Do not assume it will pause and ask.
Additional tips for getting the best results
- Use the full context window deliberately. Do not chunk inputs out of habit. 1M tokens can hold an entire codebase, an entire documentation library, or a full onboarding flow. Feed the whole thing in one go instead of slicing it and hoping the model remembers.
- Match reasoning effort to the task. The API supports
reasoning_effortatlow,high, ormax(max is the default at launch). Max reasoning on a simple task is wasted compute and time. Drop tolowfor small queries and save deep effort for long builds. - Send images via base64 or uploaded file ID, not public image links. Public image URLs are not supported on the API — this trips up developers migrating from other models.
- Give it a finish line. K3 is built for long-horizon work. Tell it what "done" looks like, then let it run instead of interrupting every 2 minutes.
What Does Kimi K3 Mean for You?
For developers and builders: Kimi K3 is the best free-to-try model for frontend work and long-context coding tasks. The 1M context window means you can paste an entire repository and ask it to build a feature, fix a bug, or generate documentation — no chunking, no retrieval pipeline. The API is cheap enough for real production use at $3/$15 per million tokens.
For small businesses: You can use K3 for free on kimi.com to draft marketing copy, analyze documents, or generate landing pages. The vision capability means it can look at screenshots of its own output and self-correct visual work. No setup or payment required to get started.
For teams with strict data rules: The open weights let you host the model on your own infrastructure and keep all data in-house. The trade-off is infrastructure cost — this is a datacenter deployment, not a laptop job.
For the open-source AI ecosystem: K3 proves that an open-weight model can match or beat closed frontier models on specific, real-world tasks. The Frontend Code Arena result (1,679 Elo, #1) came from blind human votes, not a vendor-run chart. Every open model that approaches frontier capability moves the entire field closer to commoditizing intelligence — and that benefits anyone who pays for AI.
How Does Kimi K3 Compare to the Alternatives?
| Feature | Kimi K3 | Kimi K2.6 | Claude Fable 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Developer | Moonshot AI | Moonshot AI | Anthropic | OpenAI |
| Parameters | 2.8T (MoE) | Not disclosed | Not disclosed | Not disclosed |
| Active params/token | 104B | ~32B | N/A | N/A |
| Context window | 1,048,576 tokens | 256K tokens | See Anthropic docs | See OpenAI docs |
| Open weights | Yes | Yes | No | No |
| API input $/M | $3.00 | $0.95 | See Anthropic | See OpenAI |
| API output $/M | $15.00 | $4.00 | See Anthropic | See OpenAI |
| Frontend Arena rank | #1 (1,679) | #18 | #2 (1,631) | #3 (1,618) |
| Native vision | Yes | No | Yes | Yes |
| Always-on reasoning | Yes (effort: low/high/max) | Optional | Yes | Yes |
K3 is the pick when you need maximum context (1M tokens), open weights for self-hosting, or strong frontend coding. K2.6 is the budget pick at roughly one-third the API cost. Claude Fable 5 leads on the hardest software engineering benchmarks and conversational polish but is closed and likely more expensive. The right model depends on your task, budget, and data requirements.
If you want to explore how to use Kimi K3 for specific real-world tasks, see our Kimi K3 workflow guide for business which covers four practical prompt workflows. For the financial and strategic implications of the Kimi K3 release and Moonshot's $35B valuation, see our Moonshot AI investment risk analysis.
FAQ
Q: Is Kimi K3 actually open source?
A: Kimi K3 is open-weight, meaning the trained model parameters are freely downloadable under the Kimi K3 License. This allows download, deployment, and fine-tuning. However, the training data and full training pipeline are not released, so it is not "open source" in the strict OSI definition. For most practical purposes—using, modifying, and deploying the model—open weights are sufficient.
Q: Can I run Kimi K3 on my laptop or a single GPU?
A: No. The model weights require roughly 1.4 TB of memory to load in MXFP4 format. Moonshot recommends serving on setups with 64 or more accelerators. No consumer GPU (24-32 GB) or laptop can host it. Self-hosting is a datacenter-class infrastructure project.
Q: Is the Kimi K3 API free?
A: No — the API is pay-as-you-go at $3.00 per million input tokens and $15.00 per million output tokens, with a $0.30 per million cache-hit rate. The free path is the kimi.com web chat, which includes K3 on its free tier with daily limits.
Q: What is the context window of Kimi K3 and how does that compare?
A: Kimi K3 supports 1,048,576 tokens (approximately 1 million tokens). That is 4x larger than the previous Kimi K2.6 model (256K tokens). It means you can pass an entire codebase, a full book, or a complete document library in a single prompt without chunking.
Q: Does Kimi K3 beat Claude or GPT?
A: On specific benchmarks, yes — K3 ranks #1 on the Frontend Code Arena (human preference on frontend coding) and wins on long-context agentic tasks like BrowseComp, SWE-Marathon, and Terminal-Bench. But Claude Fable 5 leads on the hardest software engineering benchmarks (FrontierSWE, DeepSWE) and overall conversational UX. Moonshot acknowledges K3 "trails Claude Fable 5 and GPT-5.6 Sol" in overall quality.
Q: Can you turn off Kimi K3's reasoning/thinking mode?
A: No — Kimi K3 always thinks. On the API, you can set reasoning_effort to low, high, or max (max is the default at launch), but you cannot fully disable thinking. For simple tasks, use low to save time and tokens.
Q: What makes Kimi K3's architecture different from GPT or Claude?
A: K3 uses three innovations: Kimi Delta Attention (KDA) for efficient long-context processing, Stable LatentMoE (896 experts, 16 active per token) for massive capacity at low inference cost, and Attention Residuals (AttnRes) for better scaling. Moonshot reports roughly 2.5x scaling efficiency improvement over Kimi K2. The result is a 2.8T model that has the running cost of a ~100B model.
Discussion
0 comments