Kimi K3: A Frontier Model You Can Measure but Not Inspect

Moonshot AI announced Kimi K3 at the World AI Conference in Shanghai on July 16: 2.8 trillion parameters, the largest open model anyone has put a name to. The benchmark results are genuinely good — fourth on Artificial Analysis’s Intelligence Index, second on a private long-horizon agent eval, first on Arena’s frontend coding board.

You cannot download it. There is no repository. The license has not been named. And the one number that would tell you what it costs to serve — active parameters per token — is missing from every official document Moonshot has published.

That combination is what makes this interesting, and it is not the story the coverage is telling. The usual arrangement with model releases is that vendors publish architecture and hide behavior: here are our parameter counts and our attention mechanism, please accept our benchmark table. Moonshot inverted it. K3’s behavior is independently measurable right now, by anyone with an API key. What K3 is remains entirely unverifiable.

What actually shipped

Nothing, yet. That is worth stating plainly because “released” is doing heavy lifting in most headlines.

As of today, K3 exists as an API and a web app. Checked directly: there is no Kimi-K3 repo on Moonshot’s HuggingFace org — the K2 family is all there, K3 is not. No MoonshotAI/Kimi-K3 on GitHub. Moonshot’s own blog commits to weights by July 27, with the technical report landing alongside them.

So “world’s largest open model” is, at this moment, a claim about a date.

Also worth fixing the terminology that press is mangling: VentureBeat’s headline calls it open-source. It isn’t. No training data or code has been promised, which disqualifies it under any OSI-style definition no matter what license arrives. Most outlets, and Moonshot itself, say open-weights — that’s the defensible term. The license is undisclosed; recent Moonshot releases (K2.5, K2.6, K2.7-Code) used a Modified MIT License, which is a decent prior and nothing more.

The number they didn’t publish

Here is what Moonshot did disclose about the architecture:

Total parameters      2.8T
Experts               896
Activated per token   16          (~1.8% of experts)
Context               1M tokens
Attention             KDA + Attention Residuals
Vision                native

And here is what they didn’t:

Active parameters per token       — not disclosed

If you are evaluating this model as an engineer, that omission is the whole ballgame. Total parameters tell you what you must store. Active parameters tell you what you must compute per token — they drive throughput, cost per million tokens, and whether serving is viable at all. It is the number in every MoE model card because it is the number people need.

For comparison, K2 published its equivalent: 1T total, 32B active, 384 routed experts plus one shared. K3 goes ~2.8× wider on totals and 4× more granular on experts (384 → 896) while cutting activation ratio hard. Those are real architectural moves. What they cost per token is unstated.

Press estimates span 40B to 100B. That is a 2.5× spread, which is another way of saying nobody knows.

A small case study in how numbers become facts

Latent Space’s AINews ran the headline “Kimi K3 2.8T-A50B” — the A50B implying 50 billion active parameters, in the standard MoE naming convention that makes it look like a spec.

That figure appears nowhere in Moonshot’s materials. The piece doesn’t source it. It reads as the newsletter’s own inference from the activation ratio — a reasonable back-of-envelope, rendered in the notation reserved for disclosed specs.

It is now propagating as a fact, because it was in a headline and headlines are what get copied. Watch for it. In two weeks “A50B” will be in a dozen comparison tables, and the weights will either confirm it or quietly not.

What’s actually been verified

This is where I had to correct my own assumption, and the correction is more interesting than the assumption was.

I’d assumed that with no weights out, nothing about K3 could be independently checked. That’s wrong. Artificial Analysis evaluated it through the API and spent $2,709.75 across 130 million output tokens doing so. That is real third-party measurement. Behavior does not require weights — only architecture does.

So the line runs somewhere more useful than open-vs-closed:

Independently verified (Artificial Analysis + Arena, via API, as of July 17 2026):

Eval	K3	Rank	Leader
AA Intelligence Index	57	#4 / 189	Claude Fable 5 — 60
GDPval-AA v2	1685 Elo	#4	Claude Fable 5 — 1760
AA-Briefcase	1548 Elo	#2	Claude Fable 5 — 1583
Arena Frontend Code	1679 (76% win rate)	#1	—
Arena Text	1486	#9	—
Output speed	62 tok/s	#90 / 189	—
Omniscience hallucination rate	51%	—	—

Rests on Moonshot’s word alone: the 2.8T count, 896 experts, 16 activated, the claimed 6.3× decode speedup at 1M context from Kimi Delta Attention, the ~25% training-efficiency gain from Attention Residuals at under 2% added cost, ~2.5× better scaling efficiency than K2. None of it is checkable until you can load the weights and count.

Two notes on that table. AA-Briefcase at #2 is the standout — it’s AA’s private eval built from four multi-week knowledge-work projects, thousands of input files, 91 tasks, scored on blind pairwise Elo. Long-horizon agentic knowledge work is the hardest thing to fake and the closest proxy to real usage. An open-weights model landing 35 Elo off the frontier there is the strongest single result in this release.

And the Frontend Code #1 versus Text Arena #9 spread is itself a finding. K3 is not uniformly frontier — it is spiky, excellent at some things and mid-pack at others. Aggregate indices hide that shape, which is a decent argument for reading the sub-boards rather than the summary number.

One caveat on all Elo figures: they are recomputed as models and pairwise comparisons are added, so they drift. I’ve seen 1685 and 1668 for GDPval-AA published on the same day by different outlets. Any Elo you cite needs an as-of date or it’s wrong by the time it’s read.

Give Moonshot credit on one thing: their own materials claim wins on roughly 7 of 35 programming and agent benchmarks and openly concede K3 trails Claude Fable 5 and GPT-5.6 Sol overall. That’s unusually candid for a vendor table. The “mostly beating Opus 4.8” framing floating around is downstream spin — don’t attribute it to them.

The memory wall

Here’s the part the parameter-count headlines obscure entirely.

Sub-2% activation means K3 is computationally cheap and memory brutal. Every token touches a sliver of the network — but you have to hold all 2.8 trillion parameters resident to know which sliver. Sparse compute doesn’t buy you sparse storage.

The only public hardware analysis I found scales Unsloth’s measured K2 build sizes by 2.8× (modemguides) — estimates, not measurements, and the author says so plainly:

Build              Est. K3 memory      K2 actual (basis)
─────────────────────────────────────────────────────────
Full precision     ~1.7TB              610GB
Dynamic 2-bit      ~950GB – 1TB        350GB
Dynamic 1.8-bit    ~650 – 700GB        240GB

Total memory across VRAM and system RAM has to meet the build size. Miss it and you’re in disk offload, which means under 1 token per second — technically running, practically not. The 24GB GPU plus 256GB RAM setup that ran K2’s smallest quant at ~10 tok/s doesn’t reach K3’s smallest likely build.

The honest answer to “can I run this”: no. Not on any machine a person owns. Even the most aggressive 1.8-bit quantization exceeds a fully-specced Mac Studio.

Which reframes what “open weights” means here. It does not mean you can run it. It means well-funded organizations can audit and self-host it. That’s a real and valuable thing — it’s just not the thing the phrase implies to most people reading the headline. (And caveat the caveat: these scalings assume K3 follows K2’s precision behavior. KDA and Stable LatentMoE could move them materially in either direction.)

The pricing story is the actual story

I think this is the most underrated fact in the release.

Kimi K3      $3.00 / M in    $15.00 / M out    ($0.30 cached in)
DeepSeek V4  ~$0.87 / M out

K3 is priced at parity with Western frontier models — roughly 17× DeepSeek V4’s output price, and the highest any Chinese lab has set. The-decoder reads this as the end of cheap Chinese AI, and I think that’s right but slightly understated.

The cheap-Chinese-model era isn’t ending in a price war. It’s ending because a Chinese lab looked at its benchmark position and decided it could charge Western prices. That’s a confidence signal, and it’s a more meaningful indicator of where things stand than any Elo score — pricing is a claim a company has to defend with revenue.

Worth sitting with the tension: K3 ranks #4 in intelligence and #127 in price, at 62 tok/s. Fourth-best and slow and not cheap. If you’re picking a model on economics rather than a leaderboard, that combination is a harder sell than the headlines suggest.

DeepSeek already did this, with more disclosure

The comparison that sharpens everything:

	Kimi K3	DeepSeek V4-Pro
Total params	2.8T	1.6T
Active params	not disclosed	49B
License	not named	MIT
Weights	promised 7/27	on HuggingFace since April 24
AA Intelligence Index	57	52

DeepSeek shipped a real MIT-licensed 1.6T model three months ago with a disclosed active-parameter count and weights you can pull today. Moonshot announced a bigger number, disclosed less, and has shipped nothing — and got substantially more coverage for it.

K3 is genuinely more capable; 57 versus 52 is not noise. The point isn’t that DeepSeek is better. It’s that “largest open model” is being graded on the announcement rather than the artifact, and the artifact is the only part that matters.

On the trend line, two credible sources disagree usefully. Epoch AI puts open behind closed by ~4 months / 8 ECI points since January 2026. AA’s index shows the gap going 13 → 6 points across twelve months, with K3 cutting it to 3. Different instruments, different answers, same direction.

The stronger signal isn’t either number. It’s that five independent open families — DeepSeek, Qwen, Kimi, GLM, Mistral — arrived at frontier-adjacent quality at roughly the same time. One lab getting there is luck. Five is structural. (Llama’s absence from every one of these conversations is its own story, and not a happy one.)

What July 27 decides

Moonshot says the technical report ships with the weights. That makes this a falsifiable release with a due date, which is more than most announcements offer. Three things to check:

Do the weights actually land, under a real open license? Modified MIT would match their recent pattern. Anything with usage restrictions makes “open” marketing.
Does the report disclose active parameters per token? If yes, K3 is what it says and we can finally compute its economics. If the number stays missing after the weights are out, people will count it themselves within a day — so the interesting case is a report that omits it anyway.
Does it disclose training compute, tokens, and chips? This is the conspicuous silence. Nothing about training has been published: not FLOPs, not tokens, not cost, not hardware. Given export controls — BIS moved to case-by-case review for AI chip exports to mainland China in January 2026 — the chips question is the newsworthy one, and its omission doesn’t read as an oversight. DeepSeek’s willingness to publish training cost is precisely what made its releases land the way they did.

Until then, the accurate framing is narrow: an API-only model, independently measured as fourth-best in the world and second on long-horizon agent work, priced like a Western frontier model, whose architecture is entirely unverified and whose weights are a promise with a date on it.

That’s a real achievement described honestly. It’s also a smaller claim than “the largest open-source model ever released,” which is what most of today’s coverage says, and which is wrong in two words out of six.

FAQ

How many parameters does Kimi K3 have? Moonshot’s own blog states 2.8 trillion total parameters, and separately markets it as the world’s first open 3T-class model. Those are not conflicting reports — they are the spec and the pitch from the same document. The number that actually governs inference cost, though, is active parameters per token, and Moonshot has not published it. K3 uses what they call Stable LatentMoE, activating 16 of 896 experts per token — under 2 percent — but the parameter count those 16 experts represent is undisclosed. Press estimates range from 40B to 100B, a 2.5x spread that amounts to nobody knowing.

Is Kimi K3 open source? No, and the distinction matters. It is open-weights, and as of July 17, 2026 not even that yet — Moonshot says weights land by July 27. Nothing was released at announcement; the model is API and web-app only right now. There is no Kimi-K3 repository on Moonshot’s HuggingFace org or GitHub. No training data or training code has been promised, which rules out open-source by OSI standards regardless of what license eventually appears. The license itself is undisclosed. Moonshot’s recent K2.5, K2.6, and K2.7-Code shipped under a Modified MIT License, which is a reasonable prior but not a commitment.

How does Kimi K3 compare to DeepSeek V4? K3 is bigger on paper and DeepSeek shipped first with more disclosure. DeepSeek V4-Pro is 1.6 trillion total parameters with 49 billion active per token, released April 24, 2026 under MIT, with weights actually on HuggingFace today. K3 claims 2.8T total but has not disclosed active parameters, has not named a license, and has not released anything. On Artificial Analysis’s Intelligence Index as of July 17, K3 scores 57 against V4-Pro’s 52. So K3 is measurably more capable — but the claim that it is the largest open model is a statement about July 27, not about today.

Can I run Kimi K3 locally? Almost certainly not, unless you own a multi-GPU workstation or a server with roughly a terabyte of RAM. Nobody outside Moonshot and its inference partners has the weights, so every figure is an estimate — the most careful public one scales Unsloth’s measured K2 build sizes by 2.8x, giving roughly 1.7TB at full precision, 950GB to 1TB at dynamic 2-bit, and 650 to 700GB at dynamic 1.8-bit. Total memory across VRAM and system RAM must meet the build size or you hit disk offload and drop below 1 token per second. Even the most aggressive quantization exceeds a maxed-out Mac Studio. The sub-2-percent activation ratio makes compute cheap while memory stays brutal, because you must hold all 2.8T parameters resident even though each token touches a sliver of them.

How much does Kimi K3 cost compared to other Chinese models? It costs the same as Western frontier models, which is the actual news. K3 is priced at $3 per million input tokens and $15 per million output, with cached input at $0.30 — roughly 17x DeepSeek V4’s output pricing and the highest any Chinese lab has set. The-decoder framed this as the end of super cheap Chinese AI. The era of undercutting on price appears to be closing not because of a price war but because a Chinese lab felt confident enough to charge parity. Artificial Analysis also measured K3 at 62 tokens per second, ranking 90th of 189 models — it is fourth in intelligence and slow.

Have Kimi K3’s benchmark scores been independently verified? Its behavior has been; its architecture has not. Artificial Analysis evaluated K3 through the API, spending $2,709.75 across 130 million output tokens — that is real third-party measurement and does not require weights. What weights would let you verify is what the model is: the 2.8T count, the 896 experts, the claimed 6.3x decode speedup from Kimi Delta Attention. Those rest on Moonshot’s word until July 27. It is an unusual inversion. Vendors normally publish architecture and hide behavior; Moonshot shipped a model you can measure but not inspect.

Is the gap between open and closed AI models closing? Yes, though credible sources size it differently. Epoch AI puts open models roughly 4 months and 8 ECI points behind closed since January 2026. Artificial Analysis’s Intelligence Index shows the gap narrowing from 13 points to 6 over twelve months, with K3 cutting it to 3. The more telling signal is not any single number but that five independent open families — DeepSeek, Qwen, Kimi, GLM, and Mistral — reached frontier-adjacent quality at roughly the same time, which suggests something structural rather than one lab getting lucky.