Model Drop: Kimi K3

Moonshot beats Opus-class with open weights

Kimi K3 is Moonshot’s new the 2.8-trillion-parameter flagship. The K2 line was Moonshot selling the frontier at a discount. K3 is Moonshot deciding it belongs at the frontier (and pricing accordingly).

Model: Kimi K3 (kimi-k3 on the Moonshot API, moonshotai/kimi-k3 on OpenRouter). Served in the Kimi apps as K3 Max and K3 Cluster Max.

Model type: Text plus native vision input, text out. Reasoning model that ships with max thinking effort only; low- and high-effort modes are promised in later updates.

Ship date: July 16, 2026. Open weights promised by July 27

Maker: Moonshot AI (Beijing)

Pricing: $3.00 per million input tokens, $15.00 per million output, $0.30 per million on cache hits, on the Moonshot API. Same $3 / $15 on OpenRouter. The full 1M context window is included at that rate, and Moonshot claims cache hit rates above 90% in coding workloads. For reference: Claude Fable 5 lists at $10 / $50, GPT-5.6 Sol at $5 / $30, and Moonshot’s own K2.7 at $0.95 / $4.00.

Available on: Kimi.com and the Kimi App (K3 Max and K3 Cluster Max for logged-in users), Kimi Work (the new desktop app for Windows and Apple silicon), Kimi Code in the terminal, the Moonshot API, and OpenRouter (single upstream provider at launch). Hugging Face weights by July 27.

Headline benchmarks: Terminal-Bench 2.1 88.3 (above Claude Fable 5’s 84.6, just under GPT-5.6 Sol’s 88.8), Program Bench 77.8 (edges both, Fable 5 at 76.8 and Sol at 77.6), GPQA-Diamond 93.5, MMMU-Pro 81.6, MathVision 94.3, DeepSWE 67.5 (trails Fable 5’s 70.0 and Sol’s 73.0). Moonshot’s own framing: overall intelligence second only to Fable 5 and Sol among everything it tested, which places K3 ahead of Opus 4.8, the Gemini line, and every open model on its internal set.

Other info: 2.8 trillion total parameters, which would make it the largest open-weight model ever released if the July 27 drop happens. New architecture rather than a K2 scale-up: Stable LatentMoE with 16 of 896 experts active per token, Kimi Delta Attention (KDA), and Attention Residuals (AttnRes). 1M-token (1,048,576) context window, four times the K2 line’s 262K. Knowledge cutoff undisclosed. License terms unpublished at launch (the K2 family ran Modified MIT). No system card; Moonshot says a technical report is coming. The announcement itself flags three limitations: sensitivity to thinking history, “excessive proactiveness” on agent tasks, and a “noticeable gap in user experience” against Fable 5 and Sol.

More details: Kimi K3 announcement

Moonshot AI dropped Kimi K3 as its first true flagship since the K2 family began iterating, and the positioning is a departure. K2 releases always positioned themselves as advantageous on pricing; near-frontier capability at a tenth of the cost. K3 pivots to a capability story with a frontier-adjacent price attached. It’s a new architecture (Stable LatentMoE, 2.8T total parameters, 16 of 896 experts active, Kimi Delta Attention), a 1M-token context window, native vision input, and a rate card of $3 in / $15 out that roughly triples K2.7 on input and nearly quadruples it on output while still coming in at 30% of Claude Fable 5’s pricing.

Moonshot did publish head-to-head numbers against Claude Fable 5 and GPT-5.6 Sol, despite performing more poorly. Terminal-Bench 2.1 at 88.3 beats Fable 5’s 84.6 and sits half a point under Sol, Program Bench at 77.8 edges both, GPQA-Diamond lands at 93.5, and DeepSWE at 67.5 trails Fable 5 by 2.5 points and Sol by 5.5. The demo reel leans on long-horizon agentic work, including autonomous GPU kernel optimization, compiler development, game and 3D content creation, and video editing, and the blog explicitly concedes a “noticeable gap in user experience” against the two closed flagships. The flagship model runs at max thinking effort only (Arena testers clocked individual tasks at 35 minutes, and BenchLM measures 62 tokens per second), and the weights are a promise so far.

And there’s no system card. Again.

What’s new

K3 is the first genuinely new Moonshot model since K2 shipped, not another post-train on the 1T MoE base. Four things separate it from both its predecessors and the frontier.

• A new architecture, not a bigger K2. Stable LatentMoE at 2.8T total parameters with 16 of 896 experts active per token, plus Kimi Delta Attention and Attention Residuals. Every K2.x release since early 2026 reused the same 1T / 32B-active skeleton so deployments could swap weights in place. K3 breaks that compatibility on purpose, and the scale jump (nearly 3x total parameters) is the largest any open-weights lab has attempted.

• A 1M-token context window at a flat price. Four times the K2 line’s 262K, matching the Gemini and Muse Spark tier, with the full window included at the standard $3 / $15 rate instead of the long-context surcharge most providers charge. Combined with the claimed 90%+ cache hit rate on coding workloads, whole-repo agentic sessions stop requiring context gymnastics.

• Benchmarks with losses printed on them. K2.7 shipped self-graded deltas over K2.6 (and I called it out for that at the time). K3 ships a table where Moonshot loses DeepSWE to both Fable 5 and Sol and says so, alongside an admitted UX gap. That’s a credibility posture no Chinese lab has taken this year, and it makes the wins (Terminal-Bench over Fable 5, Program Bench over both) considerably harder to dismiss.

• Kimi Work grows a product surface. Widgets render interactive components directly inside a chat, Dashboard gives persistent per-project views, and the whole thing ships as a desktop app on Windows and Apple silicon. Moonshot is no longer just selling an API and a terminal agent. It’s chasing the prosumer workspace the way OpenAI and Anthropic do.

How and where to use it

Where it runs, what it’s actually good for, and where you’ll regret reaching for it.

Where it’s available

• Kimi.com and the Kimi App, with K3 Max and K3 Cluster Max tiers for logged-in users

• Kimi Work on desktop (Windows and Apple silicon)

• Kimi Code for terminal and IDE agent work

• The Moonshot API via OpenAI- and Anthropic-compatible endpoints (set model to kimi-k3)

• OpenRouter at the same $3 / $15

• Open weights on Hugging Face by July 27, license terms to be confirmed

What it’s good at

• Long-horizon terminal and agent work, where Terminal-Bench 2.1 at 88.3 puts it above Claude Fable 5 (!)

• Multi-language program synthesis (Program Bench 77.8, ahead of both closed flagships)

• Whole-monorepo and document-pile workloads that actually need the 1M window

• Vision-grounded work on screenshots, charts, and documents (BenchLM ranks it #24 of 112 on grounded multimodal, with MathVision at 94.3)

• The 3D, front-end, and game-dev generation that flooded timelines during the Arena preview

• Hard science questions (GPQA-Diamond 93.5, above Fable 5)

What it’s bad at / shouldn’t be used for

• Repository-level bug fixing where the closed frontier still leads (DeepSWE 67.5 against Sol’s 73.0)

• Anything latency-sensitive, because max-only thinking effort at 62 tokens per second means single tasks can run half an hour

• High-volume boilerplate, where K3’s $15 output plus heavy reasoning burn loses the invoice math to K2.7 at $4 (keep the cheap model for cheap work)

• Local inference, because 2.8T parameters is a rack, not a workstation

• Regulated or data-sovereignty-sensitive workloads, where a Beijing-hosted API with no system card and the K2 family’s unaddressed red-team findings remains a non-starter regardless of the benchmark table

First impressions

The positives

The benchmark-to-production gap has been the K2 family’s chronic disease, so the notable thing in the Hacker News thread is people believing the numbers. HN user natrys, surveying the benchmark table:

“Generally looks like a Sol/Fable tier model, better across the board than Opus 4.8.”

And user InsideOutSanta, after actually driving it: “After using it for a few hours, I believe these benchmarks.” Day-one hands-on believers is rare, especially for Moonshot. K2.6 and K2.7 both spent their launch weeks fighting skepticism about self-reported evals.

From TestingCatalog’s roundup:

“…the level of detail, polish, and overall quality is honestly wild…”

One tester called a K3 generation “one of the best outputs I’ve ever seen from this prompt, better than many frontier models.” Interactive 3D scenes, voxel animations, and front-end builds are demo-bait workloads, sure. But blind side-by-sides where testers rank an anonymous checkpoint above Fable 5 on visual richness are exactly the kind of signal a launch blog can’t buy.

The pricing debate on HN conceded the capability point even while arguing the invoice. User Tiberium:

“…1M context, pricing is $3/$15 for 1M tokens (cache $0.3), which is extremely high for a Chinese open-weight model, but if it’s truly competitive with most of the current frontier and is only behind Fable/Sol, the pricing is justified.”

The negatives

The sharpest pushback on HN is that the price destroys the reason Moonshot models get picked in the first place. User nullbio:

“This is too expensive to be a viable model. If it were $5/1m output, it might be another story. At these prices, there’s no reason to use this over GPT 5.6.”

A max-effort-only reasoner can lose on cost even when it wins on sticker price; if Sol solves a task in 10K reasoning tokens and K3 burns 50K getting there, Sol wins. Until Moonshot ships the promised lower effort modes, every K3 call pays the maximum thinking tax.

On X, Pranesh Prakash put a number on how hollow “largest open model ever” is for the people who actually run open models:

“Of course, at 2.8T params (!!) it is waaay too large for regular consumers to run locally.”

The weights also aren’t out; they’re promised by July 27, eleven days after the API launch.

The jurisdiction and safety story hasn’t moved an inch, and the HN thread said so plainly. User austinthetaco:

“Lots of organizations are continuing to refuse to use chinese models due to security and IP concerns.”

Moonshot continues to ship without a system card, now on its most capable model ever. The independent red-team evaluation of K2.5 documented fewer CBRNE-adjacent refusals than the closed frontier, elevated compliance on disinformation requests, and political bias in Chinese-language outputs. No Moonshot release since has addressed any of it. A new architecture is not a new safety posture.

Jake’s take

Moonshot has spent all year being my favorite budget option. I’ve now written twice that the K2 family’s job was eating boilerplate while the big lads kept the architecture calls. K3 is the first Moonshot model that wants in on the architecture calls. It beats Fable 5 outright on Terminal-Bench, edges both closed flagships on Program Bench, and does it at 30% of Fable’s rate card with a 1M window.

The K2.7 drop shipped self-graded deltas and at the time I said a benchmark card that grades itself isn’t a number I can act on. K3 shows its losses, which means I can take the wins more seriously.

But at max-only thinking effort with $15 output, K3 loses the plot a bit. The HN thread’s math is right that a model spending five times the reasoning tokens loses to Sol on the invoice even at half the sticker price, and the 35-minute Arena task times show that the scenario isn’t hypothetical. It’s a slow model.

Also, the weights are a promise dated July 27, so “largest open model ever” is marketing until it’s a repo (and at 2.8T parameters, it’s a repo almost nobody can serve anyway). And Moonshot shipped its most capable model in company history with no system card, again, while the K2.5 red-team findings sit unaddressed for the fifth straight release. Bummer.

(But I’ll still be using it).