Kimi K4 Release Date: What's Confirmed vs Rumoured

• No Moonshot announcement exists. No blog post, GitHub repo, Hugging Face model card or API listing for Kimi K4 has been published anywhere.
• The rumour traces to one report. The Information reported around 29 July 2026 that Moonshot is seeking more Nvidia Blackwell-class chips for a next-generation model; DigiTimes, CryptoBriefing and others repeated it, which is amplification, not independent confirmation.
• Zero real specifications exist. No parameter count, benchmark, context window, licence or price for K4 has been published by anyone.
• The real fact underneath the rumour: Kimi K3 is a genuine 2.8-trillion-parameter, 104-billion-active-parameter model that only just finished rolling out its full weights on 27 July 2026, and is already reportedly built on roughly 20,000 Nvidia chips.
• An unresolved sanctions dispute sits directly in the way. The White House has accused Moonshot of export-control evasion and distilling Anthropic's Fable model into K3; a further chip hunt for K4 would escalate, not sidestep, that dispute.
• Nothing here should change a build decision today. Kimi K3 is shipped, priced and independently usable now; K4 is, at most, a procurement signal.

The entire public evidence base for "Kimi K4" is a single investigation by The Information, published on or around 29 July 2026 and paywalled at the source, reporting that Moonshot AI is seeking additional Nvidia Blackwell-class GPU capacity specifically to train its next-generation model. That report was then summarised by DigiTimes, picked up by CryptoBriefing, and analysed by Kingy AI, whose own piece contains an unusually direct caution worth quoting: "repetition of The Information's reporting across multiple outlets does not turn one investigation into five independent investigations."

That is the whole chain. One outlet's sourced report about chip procurement, re-published by several others, none of whom appear to have independently verified the underlying claim with Moonshot. No outlet in this chain claims to have seen a model card, a benchmark result, or a statement from Moonshot confirming a model named "K4" exists in any specific form.

As with any rumour, the story only holds together if these categories stay separate:

• Confirmed by Moonshot: nothing named "Kimi K4" exists in any official capacity. Moonshot has not acknowledged or denied the chip-procurement report.
• Reported by The Information, not independently verified by us: that Moonshot is seeking additional Blackwell-class Nvidia chip capacity for a next-generation model.
• Pure amplification: DigiTimes and CryptoBriefing's coverage, which repeats the same underlying claim rather than adding new sourcing.
• Speculation with no sourcing at all: the specific framing in YouTube video titles like "bigger than anyone expected" and "scary good", neither of which is traceable to any benchmark, model card or Moonshot statement.
• Independently verifiable today: Kimi K3's actual, published architecture, benchmarks and pricing, all from Moonshot's own GitHub repository, Hugging Face listing and API documentation.

Because there is no K4 specification to report, the only honest way to size up the rumour is against what Kimi K3 actually does, right now. K3 is a 2.8-trillion-parameter Mixture-of-Experts model with 104 billion active parameters, a 1,048,576-token (1M) context window built on Moonshot's Kimi Delta Attention and Gated Multi-head Latent Attention layers, and 896 routed experts. The API launched 16 July 2026; full weights finished publishing to Hugging Face around 26–27 July under Moonshot's own modified licence.

On independent evaluation, Artificial Analysis ranks K3 around 57 on its Intelligence Index, the first open-weight model to reach that tier, sitting behind Claude Fable 5 and GPT-5.6 Sol but ahead of Claude Opus 4.8. On LMArena's Frontend Code Arena, K3 took the top spot outright with an Elo of 1,679, a 17-place jump from its predecessor. Pricing sits at $3.00 per million input tokens on a cache miss, $0.30 on a cache hit, and $15.00 per million output tokens, a genuine step up from the ultra-cheap pricing that defined earlier Kimi releases.

That is not a small, cheap model quietly iterating. It is Moonshot's largest, most expensive, and most capable release to date, and it finished its full rollout barely a week before the K4 chip-procurement report surfaced.

Any Kimi K4 chip hunt does not happen in a vacuum. Kimi K3's own launch is already entangled in an active US-China dispute: on 22–23 July 2026, the White House publicly accused Moonshot of accessing Nvidia GB300 (Blackwell) chips via servers in Thailand to evade US export controls, and separately alleged large-scale distillation of Anthropic's Fable model into K3's training. US Treasury Secretary Scott Bessent followed with a threat of Commerce Department Entity List sanctions, which would cut Moonshot off from US chips, software and cloud infrastructure if confirmed. See our full Kimi K3 crisis timeline for the GPU capacity halt, the market reaction and the sanctions threat in detail.

Industry reporting places K3's own training run at roughly 20,000 Nvidia chips routed through Alibaba cloud infrastructure. If Moonshot is now seeking more Blackwell-class capacity for a successor, as The Information reports, that procurement effort would be happening while the Entity List threat over the previous model remains open and unresolved, not after it was settled. That is the real, verifiable context the K4 chip story sits inside, independent of whether a K4 model itself ever ships.

None of this means a Kimi K4 is unlikely eventually, only that it is unconfirmed now. Moonshot's release cadence through 2026, K2, K2.5, K2.6, the coding-focused K2.7 Code, then K3, has run roughly every six to eight weeks. A lab shipping on that schedule reaching for more training capacity within two weeks of finishing K3's rollout is consistent with its own pattern, not an anomaly.

Competitive pressure adds a second, independent reason. Alibaba's Qwen 3.8 Max launched 3 August 2026, a 2.4-trillion-parameter, 95-billion-active-parameter open model explicitly positioned to close the size gap with Moonshot and become the top-ranked Chinese model on Arena.ai's text leaderboard, though it still trails Claude Fable 5 there. DeepSeek V4 continues to undercut both on price. A lab that just spent its pricing advantage to compete on raw capability, as K3 did against K2.6, has an obvious incentive to keep pushing scale rather than cede the largest-open-model position it just claimed.

The K4 chip story lands in an unusually crowded few days. Qwen 3.8 Max shipped 3 August, DeepSeek V4 Flash received a major update the same week, and Western labs continue shipping on their own separate cadences. That density means any individual procurement rumour, this one included, should be read cautiously: a lab under visible competitive pressure has every incentive to let "we are buying more chips" leak, whether or not a specific next model is anywhere close to ready.

A Manifold Markets prediction contract, "Kimi K4 released in 2026?", priced K4 shipping in some form, open weights, chat integration, or API access, by 31 December 2026 at roughly 55% as of early August. A coin-flip-adjacent probability from a public forecasting market is itself useful evidence: it confirms informed observers treat K4 as a real possibility rather than a settled fact in either direction, and it is a meaningfully lower confidence level than the tone of "Kimi K4 Is Coming"-style video titles implies.

• Will it even be called "Kimi K4"? No source, including The Information's original report, claims to have seen a confirmed product name; Moonshot has occasionally used different naming for variants (K2.7 Code rather than K3).
• Does more chip capacity mean a bigger model, a faster one, or both? The Information's report describes procurement, not a training plan or target architecture.
• How does the sanctions dispute resolve first? Whether the US Entity List threat over K3's alleged export-control evasion escalates or is resolved will materially affect what chip access Moonshot even has for a successor.
• Does an August-adjacent timeline survive contact with reality? Every rumoured AI release covered on this site in 2026, Grok 4.6, Gemini 4, GPT-5.7, Claude Fable 5.1, has carried a stated or implied timeline that needed revision.

Teams currently building on Kimi K3 have no reason to pause: it is shipped, priced and independently benchmarked today, while K4 is, at best, a chip-procurement signal reported by one outlet. Anyone tracking the US-China AI chip dispute should watch this as an escalation vector on an already-open case rather than a separate story: further Blackwell-class procurement by Moonshot, if confirmed, would be relevant evidence in the ongoing sanctions question regardless of what model it eventually trains.

Journalists and analysts citing this story should be precise with attribution: it is accurate to say "The Information reports Moonshot is seeking more Nvidia chips", it is not accurate to say "Moonshot confirmed Kimi K4", and collapsing that distinction is exactly how a single sourced report calcifies into a treated-as-fact model name.

Strip away the video titles and the entire Kimi K4 story is one paywalled report about chip procurement, repeated by a handful of outlets, landing in the middle of an unresolved sanctions dispute over the model that came before it. The genuinely real fact underneath it, that Kimi K3 is Moonshot's largest, most capable model yet and only just finished its rollout, is a reasonable basis to expect the company keeps pushing scale. It is not evidence that a successor has been built, named, benchmarked or scheduled.

The practical position is unchanged: evaluate Kimi K3 on its own real, published merits today, and revisit this page once Moonshot, not a chip-procurement leak, actually publishes something. This article will be updated the moment that happens.

• DigiTimes summary of The Information's Moonshot chip-procurement report
• Kingy AI's analysis, including the caution against treating repetition as confirmation
• Manifold Markets: "Kimi K4 released in 2026?"
• Moonshot AI's official Kimi K3 GitHub repository
• Kimi K3 on Hugging Face
• CNBC on the White House export-control accusation