·
Moonshot AI released Kimi K3 on July 16, 2026 — a 2.8-trillion-parameter open-weight model whose self-reported benchmarks beat Claude Opus 4.8 and GPT-5.5 on most tasks, trailing only Claude Fable 5 and GPT-5.6 Sol. Kimi K3 is available through the kimi.com website and API, and Moonshot AI released the model weights for download on July 27, 2026 under a Modified MIT license — making it the largest open-weight model available for self-hosting at 2.8 trillion parameters. Independent testing has also flagged a 51% hallucination rate that Moonshot AI did not disclose alongside its benchmark charts.
Kimi K3 is Moonshot AI’s third frontier model, following Kimi K2.6, and the largest open-weight model released to date at 2.8 trillion parameters — ahead of the previous open-weight leader, DeepSeek V4 Pro, at 1.6 trillion. The model carries a 1-million-token context window and prices API access at $3 per million cache-miss input tokens, $0.30 per million cache-hit input tokens, and $15 per million output tokens — flat across the full 1-million-token context with no long-context surcharge, according to TechTimes. That pricing represents a roughly fivefold increase over Kimi K2.6’s prior API rates, signaling that Moonshot AI is positioning Kimi K3 as a premium frontier product, not a discount alternative.
What the Kimi K3 Benchmarks Show
The Kimi K3 benchmarks place it ahead of every other open-weight model on reasoning and agentic tasks, and behind only two proprietary frontier models. Kimi K3 scored 93.5% on GPQA Diamond — the strongest open-weight result recorded on that benchmark at release — and reached a 1547 Elo on Artificial Analysis’s private long-horizon knowledge-work evaluation, behind only Claude Fable 5. The verified scores across 5 published benchmarks are listed below.
| Benchmark | Kimi K3 Score | What It Measures |
|---|---|---|
| GPQA Diamond | 93.5% | Graduate-level science reasoning |
| Terminal-Bench 2.1 | 88.3% | Command-line task completion |
| BrowseComp | 91.2% | Autonomous web-browsing agents |
| MCP Atlas | 84.2% | Model Context Protocol tool use |
| Long-horizon knowledge work (Elo) | 1547 | Multi-step professional tasks |
On the Artificial Analysis composite indexes, Kimi K3 records a 57.11 Intelligence Index, a 76.24 Coding Index, and a 50.07 Agentic Index. The model outputs 62 tokens per second with a 1.99-second time-to-first-token, and completes a standard agentic task at a cost of $0.94 — similar to GPT-5.6 Sol, according to Artificial Analysis figures cited by Simon Willison. Moonshot AI reports these results itself; no independent lab has reproduced the full set.
Independent testing by TechTimes identified a 51% hallucination rate — up from 39% on Kimi K2.6 — that Moonshot AI did not include in its benchmark charts or launch announcement. Buyers should evaluate this figure alongside the GPQA headline before deploying Kimi K3 in production workflows where factual accuracy is required.
Why Kimi K3 Matters for Businesses Building AI Agents
Kimi K3’s agentic benchmarks make it a credible option for companies building AI agents on top of a hosted model. Its BrowseComp score of 91.2% measures autonomous web-browsing performance, and its MCP Atlas score of 84.2% measures how reliably the model uses external tools through the Model Context Protocol — the two capabilities that determine whether an agent can complete real multi-step work rather than answer a single prompt. At $3 and $15 per million tokens for cache-miss requests, Kimi K3 undercuts the frontier price ceiling while matching GPT-5.6 Sol on cost per task, which gives agent builders a lower-priced tier to evaluate against the best AI coding assistants for autonomous development. Cache hits drop the input cost further to $0.30 per million tokens, making repeated-context workloads significantly cheaper than cache-miss pricing suggests.
How Kimi K3 Compares to Other Open-Weight Models
Kimi K3 takes the open-weight parameter crown at 2.8 trillion parameters, ahead of DeepSeek V4 Pro at 1.6 trillion, and is the first open model from a Chinese lab to challenge United States frontier labs on agentic and reasoning benchmarks at the same time. Moonshot AI competes directly with OpenAI and Anthropic for API market share, and Kimi K3’s GPQA Diamond result of 93.5% is the strongest open-weight score on that benchmark. Buyers weighing these numbers should read the self-reported scores against independent evaluation, as explained in our guide to how to read AI coding benchmarks.
Kimi K3 does not beat Claude Fable 5 or GPT-5.6 Sol overall on Moonshot AI’s own numbers — it trails both. The “open weight” label, while now active, carries important caveats: the weights require approximately 1.4 TB of storage in MXFP4 four-bit precision, and the Modified MIT license terms for commercial use had not been fully published alongside the weights release.
Kimi K3 Open Weights: Architecture, Download Size, and License Terms
Kimi K3 uses a mixture-of-experts (MoE) architecture with 896 total experts, of which only 16 fire per token — producing approximately 50 billion active parameters per inference step. That design means per-token compute resembles a 50-billion-parameter model despite the 2.8-trillion-parameter total weight count, which is how Moonshot AI achieves frontier-level output at $0.94 per agentic task.
The full weight download requires approximately 1.4 TB of storage in MXFP4 four-bit precision, according to TECHi. Running the model locally therefore requires data-center-class hardware; consumer or small-office workstations cannot host Kimi K3 in any standard configuration. Businesses planning to self-host should confirm server requirements — RAM, NVMe capacity, and GPU interconnect — before committing to a download.
The open weights are released under a Modified MIT license. As of the weights release date, Moonshot AI had not published the specific commercial-use terms alongside the weights announcement, which leaves ambiguity about redistribution and product-building rights. Anyone planning to build a product on Kimi K3’s weights should confirm license terms before proceeding, according to TechTimes.
- Chinese AI Models Are Running US Enterprise Workloads — how open-weight models from Chinese labs entered production stacks
- GLM-5.2: Zhipu AI’s 1M-Context Open-Source Model — the release that set the current open-source context-length bar
- Tencent HY3: A 295B Open-Source MoE Model — another Chinese open-weight entrant competing on cost
The Kimi K3 open weights are now live, and the 2.8-trillion-parameter headline is real — but two numbers matter more than the GPQA score. First: 51% hallucination rate, confirmed by independent testing and omitted from Moonshot AI’s launch materials. That is not a minor caveat; it means roughly 1 in 2 factual claims from the model requires verification, which limits safe deployment to workflows with a human or automated fact-check layer. Second: the $0.94 cost per agentic task remains the most competitive hosted rate at frontier agentic capability — for teams that can build a verification layer, Kimi K3 is worth a structured pilot. The 1.4 TB download and unresolved commercial license terms make self-hosting a longer planning exercise than the “open weights” label implies; most businesses will evaluate the API before committing to infrastructure.
Related Coverage
- Claude Sonnet 5 Agentic Model Launch — the frontier line Kimi K3 measures itself against
- 5 Agentic AI Security Risks Every Team Must Know in 2026 — what to check before deploying any agent model in production
Frequently Asked Questions
What is Kimi K3?
Kimi K3 is a 2.8-trillion-parameter open-weight AI model released by Moonshot AI, a Chinese lab, on July 16, 2026. The model uses a mixture-of-experts architecture with 16 of 896 experts active per token (approximately 50 billion active parameters). It is available through the kimi.com website and API, and its weights were released for download on July 27, 2026 under a Modified MIT license.
Does Kimi K3 beat Claude Fable 5?
No. On Moonshot AI’s own self-reported benchmarks, Kimi K3 trails both Claude Fable 5 and GPT-5.6 Sol. It beats Claude Opus 4.8 and GPT-5.5 on most tasks and leads all open-weight models, but it does not top the frontier overall.
How much does Kimi K3 cost?
Kimi K3 costs $3 per million cache-miss input tokens, $0.30 per million cache-hit input tokens, and $15 per million output tokens through Moonshot AI’s API — flat across the full 1-million-token context with no long-context surcharge. That pricing represents a roughly fivefold increase over Kimi K2.6’s prior rates, according to TechTimes. Artificial Analysis measured the cost per agentic task at $0.94.
Is Kimi K3 open source?
Kimi K3 is open-weight but not fully open source. Moonshot AI released the trained weights for download on July 27, 2026 under a Modified MIT license; however, the specific commercial-use terms had not been fully published alongside the weights release. The full weight download requires approximately 1.4 TB of storage in MXFP4 four-bit precision, which requires data-center-class infrastructure to self-host.
What is Kimi K3’s hallucination rate?
Independent testing identified a 51% hallucination rate for Kimi K3 — up from 39% on Kimi K2.6. Moonshot AI did not include this figure in its launch benchmark charts. Businesses should factor this rate into any deployment that requires factual accuracy, and pair the model with a verification layer in production workflows.
·
