Kimi K3 is rolling out: Moonshot's open-weight giant gunning for Opus 4.8

1. What is confirmed, what remains unclear

Confirmed. The official announcement has landed: Moonshot presents K3 as its "Open Frontier Intelligence" — 2,800 billion parameters (mixture-of-experts architecture), a one-million-token context window, native multimodality (the model understands images without a separate module), and an always-on reasoning mode. The rollout, started in waves on July 16 through the Kimi CLI, now covers Kimi.com, Kimi Work, Kimi Code and the API. Open weights are announced for July 27, 2026. The price list is live — we break it down below.

Called in advance. We wrote it in the first version of this article: the date did not leak through an engineer, but through the store. On the Kimi Open Platform, a credit top-up promotion named "K3 launch" runs from July 15 to August 11, 2026 (10 to 30% bonus). You do not name a commercial operation after a model you are not about to serve — those credits have to be honored. The marketing department had spilled the beans before the comms department. It was right, give or take a day.

Settled since. The "2,500 to 3,000 billion parameters" rumor — which we refused to endorse as long as it rested on a single anonymous source — is closed: 2,800 billion, officially. And our architectural bet is confirmed: we wrote that linear attention was "the most likely path" to serve a one-million-token context (the team had admitted in a December 2025 AMA that it was "too costly to serve" back then). That is exactly what Kimi Delta Attention does — K3's hybrid linear attention, up to 6.3x faster decoding in million-token contexts — backed by Attention Residuals (~25% higher training efficiency at less than 2% additional cost, per Moonshot).

Still unclear. Only two things: the weights themselves (announced for July 27, not yet downloadable — exact license to be checked at release), and any independent measurement of performance. The 14 published benchmarks remain vendor figures.

2. Pricing: the end of the Kimi discount

Here is the table that should catch the eye of any executive paying an API bill — prices per million tokens:

Three readings of this table, from the most obvious to the most interesting.

One. Moonshot is changing business models. At $3 / $15, K3 costs three times K2.6's input price, almost four times on output — and lines up exactly with Claude Sonnet 5. The message is clear: Kimi no longer sells itself as the low-cost alternative, but as a frontier model priced accordingly. Which is consistent with the money invested: $500 million raised in January — earmarked, according to press reports, for "K3 development and compute capacity expansion" —, then $2 billion in May at a $20 billion valuation. A model subsidized at a loss does not pay that back.

Two. The real price is the cache hit. $0.30 per million tokens on already-seen input is ten times less than the nominal rate. And that is precisely how agents consume: re-reading the same context (instructions, knowledge base, history) a hundred times, adding only a few new tokens each turn. Combined with the one-million-token window, this pricing describes a model designed for long agentic workflows — not for chat.

Three. The trade-off narrows, but does not flip. Against Opus 4.8 ($5 / $25) and Fable 5 ($10 / $50), K3 remains 40 to 70% cheaper on output — the line item that explodes in agentic usage. But the days of switching to Kimi blindly to divide your bill by six are over. From now on, you measure: at comparable quality, the price gap justifies the test; at lower quality, K2.6 at $0.95 remains the market's best value for volume work.

3. How does it stack up against Opus 4.8?

The trade press headlines that K3 is "expected to close the gap" with Anthropic's Opus 4.8. Early user reports point the same way — some even put it ahead on long coding tasks. Our position: plausible, and not yet demonstrated. No independent benchmark has been published at the time of writing. We have seen too many models shine on social media for three days before falling back in the measurements to treat impressions as data.

The full benchmark card is now out. After three early scores relayed on launch night, Moonshot published its complete card: 14 benchmarks, "all maxed out on thinking effort" — every model measured at its maximum reasoning effort. It also corrects the early figures: 1,668 Elo on GDPval-AA v2 (not 1,687), 1,548 on AA-Briefcase (not 1,527), and GPT-5.6 Sol drops to 90.4% on BrowseComp (not 92.2) — which hands that line's first place to K3. The caveat stands: these are vendor-published figures, not yet cross-checked by an independent measurement.

If these scores hold up, they describe exactly a Fable / Sol-class model: ahead of Claude Opus 4.8 on all 14 published lines, first on 5 of them (agentic browsing, spreadsheets, automation, Program Bench, long-run coding), behind Claude Fable 5 on knowledge work and most coding lines — the first-place tally reads 7 for Fable 5, 5 for K3, 2 for GPT-5.6 Sol. In other words: the level the price announces. To be confirmed — we will update this section as soon as independent measurements land.

What is measured independently, on the other hand, is the starting point. Kimi K2.6 ranks 4th worldwide on Artificial Analysis' Intelligence Index, at 54 points versus 57 for the three closed leaders at the time of measurement, in April 2026 (Claude Opus 4.7, Gemini 3.1 Pro, GPT-5.4 — since superseded by Opus 4.8, GPT-5.6 and others; the index has not been re-run yet). The world's best open-weight model, three points from the frontier. If K3 only gained those three points, the press headline would be factually fulfilled — and the asking price justified.

One corrective, still, to avoid howling with the wolves. The benchmarks Moonshot puts forward are in-house benchmarks ("Kimi Code Bench", "Kimi Claw"). On its own results table, K2.7-Code lost 11 cells out of 12 against GPT-5.5 and Claude Opus 4.8. The gap has narrowed; it had not disappeared. Anyone telling you today that K3 "has overtaken" the closed frontier is over-reading: even by Moonshot's own card, Claude Fable 5 keeps 7 first places out of 14 — including, an irony that lends the card credibility, Moonshot's own in-house benchmark (Kimi Code Bench 2.0: Fable 5 at 76.9, K3 at 72.9). Ahead of Opus 4.8, however, is what all 14 published lines say — pending independent validation. We are testing K3 through the Kimi CLI and will publish ours right here.

4. Why it will change nothing for you

Here is the part other articles will not write, because it hurts the click.

K3 being very good will change nothing about your situation. Neither to make you switch to Moonshot, nor to keep you away from it. And it is not a geopolitical question: it is an architecture question.

Remember June 13, 2026. That day, invoking national security, the US government ordered Anthropic to suspend access to Fable 5 and Mythos 5 for all foreign nationals. Unable to verify its users' nationality in real time, Anthropic did the only thing possible: cut both models for everyone, everywhere. Two weeks of downtime. Restored on July 1st.

Ask yourself the only question that matters: if your automations had been running on that model, what would have happened on the morning of June 13?

A pipeline with the provider hard-coded would have stopped. A pipeline where the model is a parameter, backed by a fallback model, would have switched over and kept running without anyone picking up the phone. The difference is not decided on the day of the outage: it is decided on day one, and it costs one line of configuration. That is not genius, it is hygiene — but it is exactly what separates an automation from a gamble.

That is why K3's arrival leaves us cold: if your pipeline is well built, testing Kimi costs you one configuration line and an afternoon of measurement. If it is badly built, no model in the world will save you — you will just be captive to another provider, in another country. This is exactly the reversibility principle we apply in our AI automation projects.

5. Three things nobody else will tell you

Self-hosting is out of reach for an SMB

"Open weights" does not mean "on your premises". Kimi K2.5's weights weigh around 595 gigabytes. You need multi-GPU data-center hardware. With K3's official 2,800 billion parameters, it will be worse. Moonshot does not publish smaller versions (8, 32, 70 billion), despite community demand. In practice, open-weight sovereignty is consumed through a European hosting provider, not in your server room.

Calling Moonshot's API means sending your data to China

No more and no less serious than sending it to the United States — but no less either. It is a transfer that must be documented and framed under GDPR. Not an ideological debate, a contract clause. And remember the deadline coming up: the AI Act's transparency obligation (Article 50) applies from August 2, 2026 — your users must know they are interacting with an AI, whatever model sits behind it.

The best model is almost never the point

In the pipelines we deliver, the gains come from data quality, step decomposition and error recovery — not from the three index points separating 1st from 4th. A significantly cheaper model that does the job is a better engineering choice than a slightly stronger model that ruins the margin. And since K3's pricing, that trade-off also applies inside the Kimi catalog: for sorting, extraction or summarization, K2.6 at $0.95 remains unbeatable.

6. FAQ