Kimi K3 Open Weights Are Live — 2.8T Parameters, Day-0 on Together and Modal

Update — July 29, 2026: Architecture deep-dive from Sebastian Raschka (LatentMoE, NoPE, KDA, attention residuals) — Kimi K3 architecture notes. Laptop existence proof: Deltafin on M1 Max (~16 s/token).

Update — July 28, 2026: Anthropic CEO Dario Amodei directly addressed this release's policy backdrop the day after weights went live. In a July 27 post, Amodei said Anthropic "has never advocated for a ban on open-weights models" and called non-dangerous open models "a public good," while still pushing for chip export controls, anti-distillation enforcement, and mandatory safety testing. Full breakdown: Anthropic's position on open-weights models.

Update — July 27, 2026: This post supersedes the "not yet available" status in explainx.ai's Kimi K3 local desktop guide, now fully rewritten for the live release — weights are downloadable today. That guide covers confirmed hardware tiers and working vLLM serve commands, not just a prep checklist.

Want to actually run it, not just read about it? Go straight to how to run Kimi K3 locally — where the weights are, confirmed hardware tiers, and working vLLM setup. This post covers the release itself; that one covers download location, GPU requirements, and hosted alternatives if your hardware doesn't clear the bar.

Moonshot AI released free, public download weights for Kimi K3 on July 26, 2026, at roughly 7:30 PM EDT — a day ahead of the July 27 target the company had previously communicated. Confirmed specs: 2.8 trillion parameters and a 1,048,576-token (1M) context window, both documented on Moonshot's own Kimi K3 platform pages. It follows the model's initial API and consumer-app launch on July 16, 2026, which explainx.ai covered in the Kimi K3 API guide.

Cloud platforms moved fast: Together AI and Modal both announced day-0 hosted access, timed to the drop. This post separates what's confirmed — parameter count, context window, hosting partners, release date — from what's still a claim circulating in social coverage, including the "largest open-weight model ever" label and a "under 24 hours" launch-to-weights framing that doesn't hold up against Moonshot's own timeline.

TL;DR — what actually happened July 26

What Moonshot actually shipped

The confirmed facts, all traceable to Moonshot's own Kimi K3 tech blog and platform documentation:

Total parameters is not the same as active parameters per token — with 16 of 896 experts active, the real compute footprint per forward pass is a small fraction of the 2.8T headline number. That distinction matters for anyone sizing hardware: see the full architecture breakdown and confirmed hardware tiers in explainx.ai's local-inference guide.

Why "open weights" — not just an API — matters here

Kimi K3 has been available via API since July 16. The open-weights release changes what developers can actually do with it:

• Self-host and modify. Teams can fine-tune, quantize, or run K3 fully air-gapped, which a hosted API cannot offer regardless of pricing.
• Independent inspection. Open weights plus Moonshot's technical report give outside researchers their first real chance to examine MoE routing and tokenizer structure — relevant given unverified distillation chatter that circulated on Hacker News after the initial launch.
• No vendor lock-in on inference cost. Once weights are public, pricing competition shifts to whoever hosts them best — which is exactly what Together AI and Modal are now doing.
• Auditable licensing for commercial use. Moonshot's prior open model, Kimi K2.7 Code, shipped under a Modified MIT license. That was the expected default for K3, but confirm the actual LICENSE file in the Hugging Face repo before shipping K3 in a commercial product — don't assume terms carried over automatically.

This is the same distinction explainx.ai has tracked across DeepSeek V4 Pro, Apertus, and GLM-5.2: an API launch gets you access, an open-weights release gets you control.

Where to run it — hosted options beyond self-hosting

Together AI and Modal confirmed same-window hosted support rather than waiting for community re-quantization, and the hosting list has grown since:

If you just want K3 working in an app today with the least setup, start with OpenRouter or the platform you're already using (Together, Modal, Fireworks) rather than provisioning GPUs — self-hosting only pays off at real production volume. See explainx.ai's OpenRouter coverage for how its aggregator pricing works.

Pricing and exact model-tag availability vary by platform and change quickly — check each provider's live docs before committing production spend, and don't assume a platform's general "Kimi" support automatically means the specific K3 checkpoint until you confirm the model ID in their catalog.

For teams that want full local control instead of a managed platform, explainx.ai's local inference guide covers where the weights actually live, working vLLM/SGLang serve commands, and realistic hardware tiers (plan on 8x H100 80GB minimum — a single consumer GPU will not load the full model even quantized) — plus the same hosted-alternative list above with more detail on when to pick each one. The OpenCode local model guide covers wiring either a hosted or self-hosted endpoint into an agent harness.

How fast was this, really? Checking the "under 24 hours" claim

One tracked source (@teortaxestex) framed the release as "strictly under 24 hours until an open frontier model." That claim needs a timeline check before it gets repeated as fact.

Per explainx.ai's own coverage:

• July 16, 2026 — Kimi K3 API and consumer app launch (full guide)
• July 17, 2026 — Moonshot's official tech blog states open weights are targeted for July 27, 2026
• July 26, 2026, ~7:30 PM EDT — weights actually go live, roughly one day ahead of that July 27 target

That's about ten days from initial launch to open weights — not under 24 hours. What likely is true and fast: Moonshot beat its own stated deadline by close to a full day, which is a genuinely quick last-mile turnaround for a 2.8T-parameter model with a technical report attached. But "beat our own promised date by a day" and "launched to open weights in under 24 hours" are different claims, and only one of them is supported by Moonshot's own public timeline. This is a good example of why claims from social coverage need a check against primary sources before they get cited as fact — see explainx.ai's broader pattern of flagging AI benchmark and marketing claims that don't survive a timeline or methodology check.

The "largest open-weight model ever" label carries the same caveat: it's a reasonable read of the tracked 2026 open-frontier field — K3's 2.8T beats DeepSeek V4 Pro (1.6T) and MiMo V2.5 Pro (1.02T) — but it is not an independently audited superlative against every open-weight release in history. Treat it as strong marketing framing backed by real numbers, not a verified record.

The policy backdrop, briefly

This release didn't land in a vacuum. The same week, Anthropic and OpenAI were reportedly lobbying for tighter restrictions on open-source and Chinese-origin AI models, with The Information describing Anthropic as "increasingly isolated" from the rest of Silicon Valley over that stance. In the opposite direction, AMD, Vercel, and Ollama publicly signed the "Open Weights and American AI Leadership" letter — originally backed by Nvidia and Microsoft — arguing for keeping open-weight models, including Chinese ones, broadly available.

That isolation framing shifted a day later. On July 27, Dario Amodei published a direct rebuttal stating "Anthropic has never advocated for a ban on open-weights models" and calling non-dangerous open models "a public good" — while still asking for chip export controls, a crackdown on industrial-scale distillation, and mandatory pre-release safety testing for any sufficiently capable model, open or closed. He names China's open-weight labs as part of the "authoritarian AI superiority" risk he's worried about, but is explicit that the risk doesn't hinge on whether the weights are open — a closed model handed to a state security service is his stated worse case, not a free Hugging Face download like K3's.

A free, 2.8-trillion-parameter download from a Chinese lab is a concrete data point in that fight, regardless of which side it's cited by. explainx.ai is covering that policy angle in depth separately — see Anthropic's position on open-weights models, the Open Weights and American AI Leadership letter coverage, and the Little Tech Association's letter opposing a Chinese open-weight ban — this post stays focused on the release itself rather than re-litigating the policy debate.

What developers should actually do this week

1. Pin a commit hash from Moonshot's Hugging Face org before wiring K3 into anything production — don't float on main.
2. Read the LICENSE file directly rather than assuming Modified MIT carried over from K2.7.
3. Compare hosted vs self-hosted economics — Together AI and Modal remove the multi-GPU provisioning problem, but per-token hosted pricing may cost more than self-hosting at high volume. Run the math against your actual request volume.
4. Re-run your own eval suite rather than trusting benchmark claims at face value — the same discipline explainx.ai applied when Kimi K3 topped nextjs.org/evals back in July.
5. Watch for GGUF and re-quantized community ports in the days following release — full BF16/MXFP4 weights are large; smaller quantizations typically follow within a week.

Summary

Kimi K3's open weights are live as of July 26, 2026 — 2.8 trillion parameters, 1M-token context, free to download, with Together AI and Modal both offering day-0 hosted access. The jump from Moonshot's July 16 API launch to open weights took roughly ten days, not the "under 24 hours" some social coverage implied — a useful reminder to check viral claims against a project's own public timeline. The "largest open-weight model ever" tag is well-supported among 2026's tracked open-frontier releases but isn't an independently audited record. The release also lands squarely inside an active US policy fight over restricting open-weight Chinese models — one this post flags but doesn't resolve; see explainx.ai's dedicated policy coverage for that argument in full.

Related on explainx.ai

• Kimi Slides — K3 research to editable PowerPoint
• What Are LLM Parameters? Top 10 Model Sizes (July 2026)
• Kimi K3 1-bit GGUF — 1.56TB shrunk to 594GB, runs on a Mac Studio
• Kimi K3 architecture — Raschka / LatentMoE / NoPE
• Deltafin — K3 on one Apple Silicon Mac
• Using an open model feels surprisingly good — OpenCode + Kimi Modal
• Intelligence ownership — Fermisense GRPO 9B
• Opus 5 built a Homeworld-style space RTS — builder retested the same prompt on Kimi K3 first
• Anthropic's position on open-weights models: no ban, but chips, distillation, and testing
• How to run Kimi K3 locally — confirmed hardware tiers and vLLM commands
• Kimi K3 API guide — 2.8T specs, pricing, 1M context
• Kimi K3 #1 on Next.js evals and Frontend Code Arena
• Kimi K2.7 Code — the prior open-weight Moonshot model
• Kimi K3 subscription pause — Moonshot's GPU capacity crunch
• Microsoft reportedly testing Kimi K3 for Copilot and Azure
• Open Weights and American AI Leadership letter — NVIDIA, Google, Microsoft
• Little Tech Association urges no ban on Chinese open-weight AI
• "American AI is losing" — the open-weights op-ed that split Hacker News
• DeepSeek V4 Pro official release and pricing
• How to run open-source models locally in OpenCode
• Why explainx.ai supports open-source AI

Official sources: Kimi K3 tech blog · Kimi Platform docs · Together AI · Modal

Release timing, parameter count, and hosting-partner claims accurate as of July 28, 2026, based on Moonshot's own platform documentation and public statements from Together AI and Modal. Claims attributed to social coverage — including "largest open-weight model ever" and the "under 24 hours" launch framing — are flagged as unverified marketing framing, not confirmed by an independent Moonshot benchmark or audit. Verify current model card details, license terms, and hosting pricing before provisioning production workloads.