The Largest Open-Weight Model Yet — With a License Nobody Read Carefully Enough
Moonshot AI’s Kimi K3 has spent two weeks generating exactly the kind of headlines a lab hopes for: largest open-weight model ever released, 2.8 trillion parameters, a 1-million-token context window, and a price point aimed squarely at the frontier tier rather than the budget shelf. What has gotten far less attention, and turns out to matter more than any single benchmark chart, is that much of what got reported about the model’s license in the days after launch was simply wrong.
Kimi K3 first went live on July 16, 2026, through Moonshot’s hosted API only. The open weights — the actual downloadable model files that let anyone self-host, fine-tune, or build commercial products on top of K3 without paying Moonshot per token — arrived on July 26, a day ahead of the July 27 date Moonshot had originally promised. That eleven-day gap between “the model exists” and “you can actually run it yourself” is worth sitting with, because it shaped almost every early claim about what K3 is and who can use it.
- API launch: July 16, 2026
- Open weights published: July 26, 2026 (one day ahead of schedule)
- Total parameters: 2.8 trillion
- Active parameters per token: 104 billion (not the roughly 50 billion figure that circulated widely in early community write-ups)
- Context window: 1 million tokens, accepting text, image, and video input (no audio)
- Download size: approximately 1.56TB on Hugging Face (native MXFP4 checkpoint around 594GB)
- Hosted API pricing: $3 per million input tokens, $15 per million output tokens, plus roughly $0.015 per web-search call
The License Correction That Matters More Than the Specs
For roughly ten days, nearly every outlet covering Kimi K3 described its license as “Modified MIT” — the same designation Moonshot used for the K2 family beginning in July 2025. That description carried over cleanly for K2, K2.5, K2.6, and K2.7 Code. It does not carry over to K3.
When the weights actually posted on Hugging Face, the accompanying document was not labeled Modified MIT at all. Hugging Face itself categorizes the license as “other” rather than under any SPDX-recognized MIT identifier. The grant is still broad in spirit — it permits use, copying, modification, distribution, sublicensing, sale, deployment, fine-tuning, and building derivative works from the weights — but it introduces something the earlier license did not have: a specific commercial threshold aimed at companies that might compete directly with Moonshot’s own hosted API business. Reporting since the weights dropped puts that threshold at roughly $20 million a year in Model-as-a-Service revenue, above which a separate commercial agreement with Moonshot becomes necessary.
If your plan is to compete with Moonshot’s own API business, the commercial revenue threshold is the number your legal team needs to check — not the phrase “Modified MIT,” which is not the license that actually shipped with K3.
This is a meaningfully different situation from a purely open license like Apache 2.0, which carries no revenue-based gate at all. It puts K3 in a middle category that has become increasingly common among frontier-scale open-weight releases in 2026: broad enough to support research, fine-tuning, and most commercial deployment, but with guardrails specifically aimed at large-scale resellers.
Open Weights Is Not the Same as Open Source
A separate distinction worth making explicit, because it gets flattened in casual coverage: “open weight” and “open source” are not synonyms. The strict Open Source Initiative definition requires that training data, training code, and a full reproduction pipeline also be publicly available. Moonshot has never released training data or full training code for any Kimi model, K3 included. What it has released is the trained parameters themselves, along with a technical report describing the architecture — enough for outside researchers to inspect routing behavior and tokenizer structure, fine-tune the model for a specific use case, or deploy it fully air-gapped, but not enough to reconstruct how the model was built from scratch.
For most practical purposes — downloading the model, self-hosting it, adapting it for a product, deploying it commercially under the license’s terms — that distinction may not matter much. For any organization with a compliance requirement that specifically demands OSI-strict open source, it matters a great deal, and K3 does not clear that bar any more than its K2-series predecessors did.
What Kimi Delta Attention Actually Changes
Beyond the licensing story, K3 introduces genuine architectural advances. Moonshot’s technical report describes a mechanism called Kimi Delta Attention, which the company says allows for decoding roughly 6.3 times faster than standard attention approaches at this scale. A second feature, described as Attention Residuals, is credited with improving training efficiency by approximately 25% relative to K2.6, K3’s immediate predecessor.
One practical wrinkle that emerged quickly after launch: K3 shipped with a single fixed reasoning-effort level rather than the low/medium/high toggle that has become standard across competing frontier models this year. Moonshot’s own blog describes the model as running at maximum thinking effort by default at launch, with lower- and higher-effort modes promised for future updates. Independent developer testing found this made even trivial requests surprisingly expensive — one widely cited test generating a single simple SVG image reportedly cost around 25 cents — and that a short, near-empty prompt still registered roughly 95 tokens, implying an undocumented system prompt of around 85 tokens attached to every call.
Where K3 Lands Against the Competition
Independent evaluation remains thin this early, and vendor-reported numbers deserve the same grain of salt applied to every frontier launch this year. With that caveat, early testing places K3 behind Claude Fable 5 on overall capability and on deep repository comprehension, but ahead of it on some long-horizon agentic tasks, and reportedly first on LMArena’s blind Frontend Code Arena evaluation. Against GPT-5.6 Sol, K3 trails on the Artificial Analysis Intelligence Index and on single-pass coding benchmarks, but narrowly leads on a benchmark called Program Bench and by a wider margin on SWE Marathon, which measures sustained, multi-step engineering work rather than one-shot problem solving.
A Live Policy Fight Underneath the Benchmarks
K3’s release also landed in the middle of an ongoing and unresolved dispute over open-weight models originating from Chinese labs. An earlier, separate accusation from Anthropic in February alleged that Moonshot used large volumes of Claude conversation logs to train prior Kimi models — a claim Moonshot has neither confirmed nor denied on the record, and one that remains independently unverified. On the other side of the debate, hardware and infrastructure companies including AMD, Vercel, and Ollama have publicly argued for keeping open-weight models broadly available regardless of country of origin, while Anthropic and OpenAI have reportedly pushed for tighter restrictions on that category of release.
The Self-Hosting Reality: Hardware, Hosting, and the First Week
Publishing 2.8 trillion parameters of open weights does not, by itself, make a model easy to run. At roughly 1.56TB in its full form on Hugging Face — with a smaller native MXFP4 checkpoint around 594GB — K3 sits well outside what a single high-end workstation, or even most single-server GPU setups, can comfortably serve at production latency. In the days immediately following the weight release, third-party hosting providers including Together AI and Modal moved quickly to offer day-zero hosted access, effectively absorbing the multi-GPU provisioning problem for teams that want K3’s capabilities without building out the underlying infrastructure themselves.
For teams that do want to self-host, early guidance from developers working directly with the release converges on a consistent set of recommendations: pin a specific commit hash from Moonshot’s Hugging Face organization rather than tracking the main branch, since a model of this scale is likely to see incremental corrections and re-uploads in its first weeks; read the LICENSE file in the actual repository rather than assuming terms carried over unchanged from K2.7 Code, given how much confusion the license question has already generated; and treat any vendor-published benchmark score as a starting point for your own evaluation suite rather than a substitute for it, a piece of advice that applies to essentially every model covered in this release cycle, not just K3. Community-driven quantization and GGUF conversions typically lag a full open-weight release by several days to a week, and K3 has followed that pattern, with smaller, more consumer-hardware-friendly versions expected to circulate through the usual open-source channels shortly after the initial Hugging Face upload.
Pricing as a Signal of Ambition
One detail in K3’s go-to-market strategy is easy to overlook next to the parameter count and the licensing dispute, but it says something meaningful about how Moonshot sees its own position: the hosted API pricing, at $3 per million input tokens and $15 per million output tokens, is deliberately aligned with Western frontier-model pricing rather than undercutting it. That is a different competitive posture than the one many open-weight Chinese models have taken in the past, where aggressive underpricing was often the primary way to win developer attention. Pricing K3 in the same range as the models it is trying to compete with suggests Moonshot is betting on being judged as a quality peer to Fable 5 and GPT-5.6, not simply as the cheaper alternative — a bet that early independent benchmarking, still catching up to the pace of the release, has not yet definitively confirmed or denied.
None of that resolves in the span of a single model launch, but it frames why K3’s licensing terms are getting more scrutiny than a typical open-weight release might otherwise attract. At 2.8 trillion parameters, K3 is not just the largest open-weight model to date — it is the first entrant in the three-trillion-parameter class, and its terms of use will likely become a reference point the next time a lab this size decides how open “open” should actually be.